I think the industry is freaking out about open weights models in general, if not specifically DeepSeek. That is why we're now on the ~4th call for pacing the frontier from the very people who, if they wanted to pace the frontier, would simply do it rather than asking Washington to get involved.
And it's why people like Hillary Clinton have been trotted out to talk about the dangers of open weights models -- I mean, does she even know what that phrase means? (I know HRC is a controversial figure and I'm not bringing her up for that purpose; I just note that she and other prominent retired politicians are now doing the circuit on Anthropic's behalf.)
hard to say but it would last some time. china currently is fast-follow mostly via distillation. they don't have the compute resources to catch up. even if their models are more efficient it's hard to beat the folks throwing an unprecedented number of gpus at their models.
Because as always this won't be about the actual pacing, but about the details of how the regulation will be implemented: How would you control the pacing? By installing a position in the company, providing regular feedback to some government organisation. This will be something the big US players can afford and implement, while an open weight blob uploaded to huggingface or modelscope, by definition, won't have a pacing officer attached, and thus will be against the law. And the Chinese companies, of course, won't abide to US law, because why would they? The open models they provide right now are essentially gifts to the world public. If the US doesn't want them, that's their choice.
So the end result will be a protectionist regime keeping the competition out, just like with cars and solar. The local industry will have a protected market, but of course won't play a role on the global stage.
Remember: small government is only good as long as it benefits the industry.
> if they wanted to pace the frontier, would simply do it
I don't think that's a fair assessment. These companies are in a Nash equilibrium where they can't unilaterally slow down without essentially destroying their company. They also can't coordinate with each other, because that's illegal. Antitrust law generally prohibits competing companies from agreeing to restrict innovation.
What I'm trying to say is that their inaction in unilaterally slowing development is consistent with their stated beliefs and requires no other motivation.
Have you looked at this backwards, though? Meaning: have you considered what it would look like if the motivations are actually just plain old money/power, but a set of stated beliefs needed to be constructed to justify what's being sought? I think these absurd beliefs make more sense that way.
> Antitrust law generally prohibits competing companies from agreeing to restrict innovation.
The antitrust claims are a complete smokescreen. Industries can and do adopt safety standards without government intervention.
> These companies are in a Nash equilibrium where they can't unilaterally slow down without essentially destroying their company.
So what? Anthropic believes their work has a 10% chance of killing all humans. I think risking the destruction of Anthropic's business should be worth avoiding that, if that's what they believe. And with one half of the frontier duopoly gone, the other half would have no incentive to race forward. And I know there's China, but they just get all their capabilities from distilling Claude, right? So, problem solved there, too.
Sure, but does slowing down the development of new models count as "adopting safety standards"? I very much doubt it.
> Anthropic believes their work has a 10% chance of killing all humans
This ignores the other part of what they believe, which is that they are the people most likely to make a model that doesn't do that. So, in their view, letting other companies win would increase the probability of human extinction.
> with one half of the frontier duopoly gone, the other half would have no incentive to race forward
I don't see how this could possibly be true, with at least half a dozen companies being just months behind what the frontier labs are releasing.
> Sure, but does slowing down the development of new models count as "adopting safety standards"? I very much doubt it.
I mean, the point (if the claim is to be believed) isn't just to "slow down the development", it's to take more time during development to properly assess the risks the models pose, develop methodologies to reduce that risk, and standardize that across companies. I doubt those wouldn't count, especially in the eyes of regulators of an administration calling for that same slow down.
> Sure, but does slowing down the development of new models count as "adopting safety standards"?
Of course! What slows down the development is the adoption of specific safety conditions the companies draw up. They can just do that, and it will hold up in court.
> in their view, letting other companies win would increase the probability of human extinction
Yes, I've heard: "We must be in charge even if we end up killing everyone in the process." I personally think that proposition is invalid, but we're all entitled to our opinions.
People used to append IANAL to such statements :-)
> I personally think that proposition is invalid
I agree, but that's meaningless in this context. I was responding to the claim that "if they wanted to pace the frontier, would simply do it", which is false, given what they actually believe.
> And I know there's China, but they just get all their capabilities from distilling Claude, right?
I hope this is a joke. But most of the LLM research and inventions come from China? The best papers are from DeepSeek? You either get the data by stealing from humans or distilling from bigger models?
If you look at their actions, they're not the actions of someone who both A) believes their technology to be an existential threat and B) doesn't want humanity to go extinct
If they truly believed both of those things, they would just shut down their companies, because being a billionaire is entirely pointless if you're dead.
I can only conclude that they don't believe both A and B. I'm gonna assume they believe B because actually wanting to exterminate humanity is too comically supervillain esque even for Altman. Therefore they must not believe A. And it makes sense. If they believe A, why are they so incredibly sloppy about security? Why did they outsource part of it to some external firm instead of leveraging their own expertise on the technology? I'm not buying it.
Instead, I think they believe C) that AI will create enormous economic value, D) that value will be distributed across the whole economy by making everyone more productive. Assuming C and D, you get E) for them to capture this value, they must maintain proprietary control of the technology in order to be able to charge everyone else for the privilege of using it.
Assuming they believe E, open models are an existential threat, not to humanity, but to OpenAI and Anthropic.
Their solution: make everyone else believe A, in order to achieve regulatory capture and somehow stop open models from advancing by banning their development, or something. This is a hail mary pass. I can just about imagine them achieving this within the US and maybe even Europe, but China? That ship has sailed.
My problem with the “making everyone more productive” is that it’s a feel-good propaganda for the individual. Most capitalist see people as an expensive cost to be eliminated. Thus all the company shares going up when they announce layoffs.
The market can only absorb so much new products, so if productivity increase, companies will reduce headcount as much as possible to increase margin for the same income, not grow their product or production to make use of their staff.
(See any wage/productivity graph)
Government will also likely follow the same path of reducing headcount instead of producing better / faster outcome for their citizens (except the internal surveillance apparatus. The one never shrinks).
Any single model at this point can do the basic type of CRUD coding most people were doing for the past 20 years.
Even a 4 bit Qwen model running locally beats me manually putting React components together by hand. But even that is too slow so we've all started using paid models in one form or another.
So I would urge these folks to calm themselves and realize Oracle made a lot of money selling managed RDBMS to people who could have easily just downloaded MySQL.
I saw it first on a recent Anthropic post, followed by a similar comment thread. There’s some sloppy techbro poetry to it, like when we used to call marketing people growth hackers a decade ago.
I bet she knows fine what the phrase means. Doesn't mean she's beyond protecting the business interests of people who fund her political goals of course but she's always struck me as a woman who knows her brief.
> I think the industry is freaking out about open weights models in general, if not specifically DeepSeek. That is why we're now on the ~4th call for pacing the frontier from the very people who, if they wanted to pace the frontier, would simply do it rather than asking Washington to get involved.
This makes no sense. Pacing the frontier gives open models the time to catch up and reach parity.
The reason people aren’t freaking out is because most people are using heavily subsidized subscriptions.
I tried the cheapest provider on openrouter and burned through $50 in a few days. Quality was ok, seems slightly above Luna quality perhaps? But that $50 is 1/4 of my codex subscription where I could have burned that many tokens or more using Astra within my weekly reset.
This won’t last forever but as long as the frontier labs are subsidizing this heavily the open models won’t matter.
With all due respect (not that I feel like much is due after that response) I am not really sure you know what I am talking about. I'm talking about inference providers, like inference.net, Fireworks, Coreweave, Digital Ocean, etc, to use DeepSeek 4.1 Flash. They didn't create the model, they just are charging you to run inference tasks. That is a different story.
DeepSeek themselves are honest about the fact that they train on inputs by default. You won't hit DeepSeek if you use OpenRouter with ZDR enabled.
No, it’s outrageously profitable above x% utilization without stealing any prompts. Provider economics still pretty good. Acquiring hardware is the current limiter.
Yep. And the model is practically unbounded in its knowledge of math. So people should keep that in mind when they read anything about its level of intelligence.
I am curious how you managed to spend that much on Deepseek via OpenRouter. I loaded $100 back in July while using v4-flash or whatever the cheap good model was at the time, and have upgraded as the new ones came out from Deepseek. I still have $16 and some of that spend also goes towards the AI usage from my customers (the context they need to load in is quite large too).
And I am using the Claude Code harness with DS as the endpoint. And I use it ~5-8hrs a day to do my coding.
Likely the user doesn't know what they're doing or has extermely bad workflows. They're prob not managing their cache, and dont use compaction.. Letting context get to 500k and invalidating their cache every 10 tool calls because they have no providor fallback settings.
I was running deepseek v4.1 pretty much non stop during work hours, with heavy tool/mcp usage and finding it very difficult to spend more than $75 in a month.
Also the cheapest providers on Openroutrr can often have terrible cache hit %, short TTLs resulting in their effective price being much more expensive than people realize. 75% cache pretty much destroys any savings from a super cheap token perspective.
It's wild how different usage patterns are between users.
I have seen the Cursor leaderboard on my company and the vibe coders consume about 5x more tokens than the developers. They and other office workers also have Claude and their limits are often over around Wednesday.
People are using millions of tokens to do very simple HTML reports. I have seen someone asking the LLM to download the entire data into the context and asking it to sort.
Now that Microsoft allows you to do /cost for individual tasks (or whatever they call them today). So I tasked Sol, Astra and Fable in cowork with exactly the same vibe coding task on the exact same zip file containing a code project I needed an update for. Astra used 20x and Fable used 15x of what Sol did.
Fable changed a lot of things I had explicitly told it not to change. Arguably a lot of them would've been correct if you didn't work in a place where abstractions are directly against the core principles, but what it produced was basically unusable. I'm not sure if Sol or Astra did best, they produced rather similar code outputs. Astra's was better, but Sol didn't do so bad. It forgot to clean up a few places after it's refactor and it made two bugs I had to correct but other than that it was fine. Astra on the flip-side might have produced code that didn't need changes but it also rewrote every piece of documentation so that it became horrible.
As far as the "experiment" goes, it just shows you that the credit consumption is basically pure magic. You'd think that the Microsoft AI admin tools and the Agent365 FOMO DLC license they sell might give you some sort of reporting, but it doesn't. What you can see is how many tokens a user consumes and the total number of tasks they've initiated as well as whatever running agents they have. You can't see what models they use or which tasks are expensive, which makes it very hard to help them. Early on we had an employee who hit their limit in an hour, and it turned out they had basically uploaded a lot of information and run it in a single long task that kept going over it again and again. We told them it might be a good idea to only give it what it needed and to create more tasks, and even though it's been three months, they have yet to consume as many credits as they did that first hour.
But that's how you support and track it. You see a user spend a lot, then you go to their computer and now that you can actually do the /cost thing, you go through their tasks and try and figure out where they're spending money...
It's obviously improving. A month ago /cost wasn't there and they just released a new dashboard for cowork, but it's still black magic that is impossible to govern.
It's because this is how AI has been sold to everyone - just ask, and it will do it.
The better pattern is to let it code the app and then you can use the app to target your data. So you only pay for it once, plus it's deterministic. But yeah, it requires setting up an environment, etc. It becomes "maintenance".
This reminds me of when we gave clients the ability to build their own Power BI dashboards. The users would end up doing a full table dump multiple times for the same table in their reports. Requiring 16GB of ram on the server and maxing out the database every time the report was refreshed.
We ended up having to hire a full time employee to fix the performance of client built reports.
I've seen the same problem with Snowflake integrated with Claude.
People run a stupid amount of expensive queries that end up costing way too much because they're asking Claude the wrong query.
Not to mention people running wrong queries, using the result as gospel, and then the result has to be sent to a data analyst to be reverse-engineered so the numbers make sense.
You can easily burn through lots and lots of money on DeepSeek, if you do eg large scale code reviews.
Eg I've used Sashiko locally for Linux kernel code reviews before sending out my contributions out to the world. Sashiko is a great system, but it can burn through tokens like there's no tomorrow.
I do some pretty insane agentic work running constantly. I’m currently burning through multiple $200 accounts every week. I regularly spend $10-$20k worth of tokens a month. Most of it has been going to my experimental c++ compiler project
I've heard that certain inference providers may have different quality of caching implementations, so even if the listed numbers are as you say, the practical cache hit % you get might be significantly different/incur significantly different costs.
When 98.5% of my requests are cache hits (according to Pi for the last week), the cache miss price isn’t that important to me, and $0.003-0.006 per 1M input tokens is shockingly cheap.
It’s also the major difference between using DeepSeek directly vs other providers also serving it, though I have not looked lately: it’s possible other providers have matched its cache hit pricing better?
It will of course depend on what you’re doing with it, but right now my session at work has a 99.8% cache hit rate, and I’ve been running this session for hours with 23M tokens read and 713K tokens written (Opus 5.5 in this case though)
There’s a big difference in speed & quality between using DeepSeek API directly with DSH vs. DeepSeek in Opencode Go with Opencode CLI. Can’t tell if it’s the provider or the harness - but worth to give it a try.
All these products, western or Chinese, are built on a hell of a lot of running rough-shod over licensing or IP laws in general.
I'm writing a program I personally need, but I would be happy if there existed something like it already, if someone else vibecoded a better version of it than mine, or if DS got better at vibing this kind of thing.
Why does liking a product mean you have to give them all of your data? People are so outraged at LG because they make the best TVs and people wanted their expensive product, yet some MBA convinced them they could make more money by spying on your entire household all the time.
It actually was awesome in the early Facebook days where you could have your entire phone contacts and other apps filled out with a profile picture and Birthday by connecting them together. But that relationship has been completely abused, privacy has been invaded, and my data has been sold to multiple companies.
The goal going forward is to keep that data private. If your company can't survive without it then I hope your company goes out of business
This isn't true at least on the API. If you read their privacy policy you'll see the training clause is scoped specifically to the consumer terms i.e. for the chat product. No such clause exists for the API service, and it would absolutely be required under Chinese law if it was taking place.
Contrary to popular belief, DeepSeek really aren't interested in your prompts.
It’s a pretty common requirement in the enterprise world. If you’re processing data for enterprise customers, it’s a lot easier to retain nothing than to deal with all the compliance issues that arise if you’re retaining data.
I have ZDR enforced and see only compatible models and providers, yet am able to use it. DeepSeek as a provider may not be ZDR, but the models are available from ZDR and no training providers on EU/US servers.
Just pin your config to a single provider, or several providers with the params `order` and `allow_fallbacks: false`. I regularly get ~98-99% cache hit rates with OpenCode. And some providers are much faster than DeepSeek; I was getting 200-300 tokens/second the other day with Together as my provider.
It's regrettable that OpenRouter doesn't even try to pin you to a single provider per session, but once you know about it, it's a problem that's easily solved.
Yes. You have to find the provider with pricing that suits your usage.
I am having 98% my input in cache, so using Coralbricks makes sense due to them giving cache reads for free — you only pay for writes. I spend maybe 5-10 dollars a day and my agents basically work day and night implementing things for me.
If your tasks are write-heavy, find a provider with cheaper output.
If you build a customer-facing app, pay a bit extra for 400+ tok/s e.g. on Lithos.
We have entered the era of building enterprise scale projects for your own personal use. I can do things it took teams years to build as a hobby project over a weekend.
the very notion that everything has to be 'impressive' to you is ridiculous. You still don't understand the era of personal software and everything has to be a windows replacement or it wasn't worth building to you?
I'm making my own azure blob storage explorer, notes mac/android, db client etc.. its not about being 'impressive' but being personally suited to individual needs.
He said most impressive and useful meaning there is no absolute cutoff. The only reason why you would be offended is because you have built nothing at all and that's you telling on yourself. Not to mention the question wasn't even aimed at you.
No, that's not what happened in this thread at all.
There's a recurring pattern on HN where any time someone talks about knocking dozens of personal projects off their list - things that almost certainly would never have actually been addressed in the finite span of a normal life, the way things go - and you AI doomers show up and demand receipts as though that's a total reasonable and definitely not obnoxious request.
It's like if you tell someone that you love your partner and they demand to sit in the cuck chair or else you're obviously lying. I keep hoping people will move past this "prove that you're actually productive" reflex, but it just keeps happening in basically every AI thread.
In reality there are many reasons not to list out projects that you've worked on with LLMs, and while "none of your damn business" is always going to be at the top, the simple truth is that I want my products and projects to be judged by what they do and how well they work, not by how they were made.
So much of that "took years" is because they didn't know exactly what the "years from now" state they were building in advance. Another huge chunk is because they started getting customers and had to respond to customer needs, demands, scale, bugfixes, preserve uptime, etc.
And even then, "enterprise" was often a dirty word in these circles. The over-engineered would-be-swiss-army-knife vendor that was mediocre-for-everyone but excellent for nobody.
I have built many tools in the recent past for myself. None of them need to be "enterprise scale." Most of them would be worse for it because the agent output suffers when the pile gets deeper and it's just adding more piles on top.
>you can't think of anything to unleash some agents on within the entire digital world at any given time?
It has to be worth it though right? Like I could spend some money and have agents build me my own Photoshop maybe (maybe?) But it would definitely be much worse to use than actual Photoshop. Then I have to have the continued interest to keep improving it which probably won't happen because the next shiny thing will grab my attention. So it all just seems like a bunch of kids that have been given a seemingly endless supply of free candy and they are going fucking nuts like chipmunks with ADHD on crack. Building all this shit that is absolutely meaningless. I realize I've gone on a rant but I'll keep going. I strongly suspect (with no evidence whatsoever) that the people who are churning slop apps out at breakneck speed have never been to an art museum. There. I said it. You've all got no taste. You wouldn't know a quality product if it hit you in the face. I'll leave with this thought- if apple didn't exist, would they ever exist now we have LLMs? I say no, because the age of good taste and refined design and original thoughts is gone forever now that we have Claude and chatgpt and agents.
I think you're underestimating how things used to be - you could go into any office, any closet, any coffee shop and find a shit-ton of half-baked, crazy-genius, kick-ass, retarded ideas and projects lying around everywhere in the world: filing systems, carpet organizing systems, outlines for film scripts, unsent letters to loved ones, etc etc etc. Whole worlds everywhere you look. And other people chipping in their two-cents worth, adding a few new filing cabinets, an idea for a film sequel, a new way to think about a different carpet, a notebook system to organize someone else's unsent letters... All this slop eventually thrown into nasty, fetid garbage dumps, forgotten.
> All this slop eventually thrown into nasty, fetid garbage dumps, forgotten.
Isn't that how it is supposed to be?
The lower the friction, the lower the signal:noise ratio.
It doesn't matter if 1 out of every 100k slop projects is actually a humdinger, how on earth will you ever find it?
The value of a project is the commitment to to it by people. Slop projects indicates a commitment in the low to none range.
So, yeah, that AI-booster who "created" (I use that word loosely) 7x Adobe replacements in a week (none of which actually work, but he'll get there eventually, I supposed) will successfully edge out the person who carefully and thoughtfully created a Photoshop replacement over six months of user feedback.
TBH, the only way to start a software business now is in stealth mode.
You have all the struggles for the price of Anthropic / cursor subscription. I use the first one I code large chunks some PR are 50k LOC and I have at least 2-3 like this a week . It’s a greenfield project .
I still have quotas left I use it for home things build 3d model of my renovation projects, alerts for shopping list etc . And yeah I use cutting edge of cutting edge of models that saves me time and money , only discount monitor saved me ~$2k on my renovation project
PR is just an entity to review /do some other LLM processes . Human part check other models review PR's, tests all kind of , security etc. Also human is to checks docs, specs in the pr, db migrations if any , some of the tests related to the PR. We stopped reading the code after opus 4.6. Sometimes for very core parts i skim through files just to make sure if the changes were correct.
I agree. I code a lot, a lot! And maybe my code is shitty, but yeah, I burn a lot of tokens, and I couldn’t do it without Chinese models. I’m just a random dev in the middle of nowhere. And I don’t feel like I’m missing out on anything with my setup at all.
But aren't you developing bad habits and learning patterns that won't work long term? Or do you think things will get cheap enough that you will be able to keep going with your current patterns post-subsidies?
> But aren't you developing bad habits and learning patterns that won't work long term?
2 reasons - there's an advantage now, use it. 2nd the frontier providers, this is the "early cheap days" like when uber was initially cheap to compete vs standard cabs. they want you to become hooked and boy are we hooked.
hooked to ai coding, but not tied to any particular model. if they decide to bump prices up, I can easily switch to a cheaper chinese model on openrouter.
Yeah it makes sense, but ultimately the only thing that creates any form of lock in is the chat history and memories, and that isn’t super important, it’s not a real network effect like a social media app or a taxi app.
Having a better model is the only real moat, without that inference is a commodity
I expect by that point we'll have local models that can do a decent job, I would guess give it a decade and we'll be running custom accelerators that are smarter than current frontier models.
In the same way that only supercomputers used to have multiple processors and caches but it's now standard.
I use the frontier openai/anthropic models at work but exclusively open weight models (on cloud/hosted inference) for personal stuff and I think about it like this; 1) I don't see any reason GLM and DeepSeek won't eventually be as good as Claude, it's just a matter of time and 2) the open models are well and truly capable enough for most of what I'd want to do. I don't need nor want an LLM chewing away on a horrible enterprise spaghetti codebase, my employers can pay for that privilege.
Long term, we will see what happens and adapt. At worst we all go back coding by hand. Meanwhile what can I do, tell my customers that I'm raising my fee because I have to pay for token? The Claude Pro $20 plan is good enough for me and even in auto mode I never had to wait for the 5 hours reset.
Compared to what a lot of companies spend on software for chip and electronics design (we're talking about $10k-200k/seat per year), AI coding assistants have a long way to go in cost before companies won't be willing to pay for them. Companies pay a fortune for software when it enables their engineers to be productive.
For my company, I'd honestly pay $4-8k/month for Claude if I had to (it would be painful, and I'd try to get cheaper options to work first). I know some enterprise Claude users are paying that much now since they have to pay for API tokens. I am certain it's at least a 2X productivity booster for our work. Compared to the cost of hiring another developer, it's well worth it.
If they stop subsidising Claude Code for the pro/max users, there will be a lot of people priced out of it, especially the casual developer. But I don't see it going away for commercial use, even with a large price increase.
Old coding is done, as a workflow in teams. It’s the top down executive pressure of being non competitive as a company, and the bottom up pressure of human laziness
Show me people handwriting code à la NASA
And I mean we as coders have been trying to do this workflow for a while, I personally would refuse to code without IntelliJ magic complete
For this workflow, there’s no going back. What’s hard to imagine is AI taking over the other workflows we predict it will; Customer service AI sucks ass for me as a customer, et cetera
> For my company, I'd honestly pay $4-8k/month for Claude if I had to (it would be painful, and I'd try to get cheaper options to work first). I know some enterprise Claude users are paying that much now since they have to pay for API tokens. I am certain it's at least a 2X productivity booster for our work. Compared to the cost of hiring another developer, it's well worth it.
That enterprise cost you're willing to pay is correlated to how much developers will work for. When driving an agent, almost anyone can do it (almost no skills required).
If devs cost $1k/m, enterprises are not going to be willing to pay $4k/m for Claude.
What I am saying is, there's an equilibrium that will be reached; the price of the human driver and the AI worker will approach each other.
Where they stabilise, I still don't know, but I'd be very surprised if, in any field (not just dev), the human gets paid multiples more than the agent they are driving, as the agents get more capable.
You should basically never pay API prices, they are always several times higher than subscriptions.
There are several open weight subscription providers. OpenCode Go used to be good but now it's complete shit. Charm Hyper is really great and the best value. Other subscriptions have a more limited model selection or provide less value but are still decent.
I freak out since months for Z.ai lite subscription, I use glm-5.3-flash every day for a ludicrous 8.5USD/month and it's as good as DS 4.1 flash, if not better.
Almost exact same experience here, but I'm using Opencode's $10/month sub. It's perma set to DS 4.1 flash and I have anywhere from 3-5 agents going at a time. Never once hit a cap of any sort. I have absolutely no idea why people would be paying $200/mo when you can get perfectly good AI for $10 from multiple places
Software has rarely been the moat. Or file formats. You have always been able to reverse engineer them. The problem, always, has been network effects.
I can build an entire, fairly useful, spreadsheet app over a weekend. But can I send my "expenses.cells" files to my accountant? Will it work with the Excel/Google docs he uses?
AI can build or reverse engineer anything as long as you are motivated enough to do it.
One example is Affinity 3 released for free. But it's a huge pain in the ass because all the guides for how to do things are for Photoshop or Affinity 2.
Your vibecoded app won't have years of reddit posts showing how to do things. This also seems to be where LLMs are the weakest at giving advice, they hallucinate 80% of the time I ask them how to do something in Affinity, giving buttons and menus that simply don't exist.
This is a knowledge problem. You can fix it by pointing the model to documentation (if it exists). Otherwise the model will give you the next best guess
Yea if i use opus 5.5 in api through openrouter and pi agent harness I will easily burn 50-100$ a day (and with fable 5.1 i could burn 200$ easily). Whereas i have now been using a claude code subscription for 2 weeks using 5.5 at all times and have never hit a limit. I often run 6+ agent sessions at once.
I do think its important long term to not be reliant on these companies as you don't have control over the system prompts, the thinking tokens, and once the subsidization stops or the company is public they will be required to start making money and thus raise prices.
But models may get more intelligent and cheaper once that time comes so it may be a non issue.
> Yea if i use opus 5.5 in api through openrouter and pi agent harness I will easily burn 50-100$ a day (and with fable 5.1 i could burn 200$ easily).
Checked yesterday, for that day alone I had used $168 worth on my $20 subscription in Claude Code. I still had plenty of weekly use left. Seems like subscriptions are discounted at a 1:10 rate?
I dont think itll be an issue. Opus 5.5 now is way more than enough for me and open weight models will reach that level by the time subsidization stops
It's enough for you now, but I feel like part of the mythology of our future is that we'll be continued to be employed because we'll be working on more complex problems, with smarter LLMs at our side.
I’d argue most “people” haven’t actually needed more in a very long time other than to keep the same old software running. The requirements bloat of operating systems, browsers, and majority of software isn’t really a generalized “people” thing; more so the state of the industry being a form of inertial bloat.
FYI you can modify the system prompt using mitmproxy. Just ask your agent to walk you through it. Anthropic system prompts are gnarly and geared towards the lowest common denominator.
there is literally a --system-prompt flag for claude code. In my mitmproxy experiments using that flag appended to the existing system prompt rather than replacing it. So I had to create a little helper to strip the system prompt sent over the wire and add my own.
Use --system-prompt-file, it will replace the whole system prompt (https://code.claude.com/docs/en/cli-reference ). You'll have to use a shell alias or function or something to always append this flag when calling Claude Code but you don't need any MitM shenanigans. Then use "/context all" to see what else is sent (here I would recommend MitM'ing since Claude Code won't show the exact tools and text), there are a lot of tools no one needs and they are bloating the context, you can deny these in the settings.json (there is also a list here: https://code.claude.com/docs/en/tools-reference ). Also set "disableClaudeAiConnectors" to false to remove even more bloat.
I tried both --system-prompt and --system-prompt-file. they both appended when i tried about 6 months ago and watched the traffic. Yes I cleanup all those tools etc.
I got insane amounts of Anthropic and OpenAI credits given to me for free for my startup, and I have not touched them.
I get privacy, freedom, and no rate limits with the GPUs I racked locally, and those are features I would never give up even if the surveillance capitalism labs paid -me- to use their models.
How many consumers are there like me? Probably not many, but once local inference hardware is plug and play, I bet the tides shift pretty quick. Also weights-on-silicon will serve the needs of most consumers locally with more speed than any GPU could deliver for a fraction of the cost.
Most people will be doing inference in their pocket or a wearable in 5 years and the giant datacenters will be like AWS, sold to only big organizations that need to auto-scale capacity of custom models on demand.
The industry surely knows this and the subsidized inference is just marketing to generate so much buzz and demand such that the tiny fraction of the market they will be able to keep in the end is big enough that they do not collapse under all the debt.
OpenAI and Anthropic will be Dell and IBM in 10 years if they survive at all.
This. At this point I don't really care about other models because max subscription are super cheap (relatively speaking) and I don't hit my limits. Even if the frontier models are only 5% better I might as well just use the best thing available if the price is reasonable.
Once the subsidization ends and cost becomes significant I will take a serious look around for the best value models and switch off the expensive providers, but that time hasn't come yet.
> Once the subsidization ends and cost becomes significant I will take a serious look around for the best value models and switch off the expensive providers, but that time hasn't come yet.
There's a reason the labs in the US frontier oligopoly are using “safety” to lobby for antitrust exemptions for mutual coordination as well as anticompetitive regulation.
I have a subscription at work, and still I find myself wishing I could use a fast Chinese model. Something wired up to really fast inference - that rapidity of feedback is a feature in itself.
4.1 Flash seems to be in that sweet spot of very decent, really fast and really cheap. Even omitting the cost, it’s still compelling for staying in flow.
People also just do different work. Opus 5.5 is a really damn good model that's even better than Astra/Fable/Sol IME and I feel a huge difference in my work.
I've been running automated research tasks for life sciences companies, and the speed in which tokens are burnt is scary. Especially when you get into a complex knowledge space and require a subwgent to reason through each possibility, token usage grows quadratically not linearly as complexity increases...
> This won’t last forever but as long as the frontier labs are subsidizing this heavily the open models won’t matter.
It's amazing how new we all perceive AI to be, and yet how old the tricks that the big players use. Their job is to just suck the oxygen out of the room as long as they have the money to do it.
Hard to enforce when you’re dealing with private companies with “creative” accounting - who’s to say what the actual cost of inference is for OpenAI or Anthropic?
Possibly they don’t even really know themselves at this point, although obviously is it significantly higher than the consumer subscription price
It's glorious isn't it. We get free work done through subsidies. At the same time, this is what threatens my job, and the money for the subsidy is basically my own invested pensions.
> The reason people aren’t freaking out is because most people are using heavily subsidized subscriptions.
My understanding is that enterprise plans don't offer those subscriptions, so they end up paying for API prices and models like these directly impact that revenue stream.
I think it's fairly likely medium to large corporations don't pay anywhere close to the listed API pricings as they get deals through existing partnerships with the big cloud providers.
MS's Copilot subsidy ended months back. The large OpenAI subsidy has just been cut in half. Who knows when Anthropic ended theirs because I don't ever remember it being great value.
What the provider actually sells to you is GPU time / load (oversimplifying). Both tokens and subscriptions are just pretty arbitrary ways to price it, almost unrelated to the actual cost of running the model.
> I tried the cheapest provider on openrouter and burned through $50 in a few days.
I wish I could observe how some of us are using these tools.
I still struggle to spend $50 in tokens per month, and I exclusively use prepaid API tokens. This is in support of personal projects and two clients. There are billing cycles where I might spend upward of $400, but this is maybe once a year. This is offset by months like August wherein I spent $12 in tokens.
The other advantage with prepaid is that it handles the other direction much better. I don't even know what a quota limit feels like. Being blocked for hours is way more expensive to me and my clients than even $500/m. Losing an entire business day over this wouldn't work out.
I've managed to convince some others to try the same thing. $200/m flat fee is a pretty extreme constant expense if you can be more clever on average.
I think a lot of people are getting pushed around by FOMO effects into spending money on pointless subsidized tokens and have (valid) fears that if they don't maintain the same apparent economic leverage as their peers that they will be left behind. This isn't actually the case, much like lines of code are a really poor indicator for the quality or productivity over a codebase.
Which models do you primarily use, and can you very roughly list your process? Agentic coding in VSCode with tons of MCPs or... something else? Do you include lots of images or have large codebases? Which agentic harness are you using?
I also find that it's easy to spend like that, but also easy not to with little impact on productivity. At the current moment I'm stuck with rider + copilot (not ideal), but e.g. using GPT 6.1 luna is really, really cheap, and lots of tasks are quickly and decently dealt with even at lower reasoning levels, (added bonus of having low latency). And that model is so cheap, I can't see a hitting 1500$ at api prices realistically - not even close. But it also depends on the harness and codebase.
I use opus primarily, on a mix of pure coding tasks, and log parsing / incident investigation.
I don't have the mental capacity to do a lot of context switching between active work streams, so I'm not doing stuff like leaving a big agent workflow running while doing other things.
I think it depends on how many threads you have running at the same time. I have Claude writing a compiler in one window, a ui framework for the language in another, an application using the installed versions of compiler and frameworks in another, and a ui designer (an Interface Builder lookalike) in another keeping up with the framework.
Each of these has a file it listens to in ~/tmp/<name>.io and whenever one needs something from the other, they message each other via that file. Tasks can bounce back and forth as issues are resolved and tested. At the same time, I keep each busy with a list of tasks
I can fairly easily run out of my $200/month subs every week, if I let Fable be the default model. With Opus it’s less likely. If and when I do, I just have an alias ‘claude.ds’ which fires up Deepseek instead, and burns through far less money, though I don’t think it’s as good at solving problems, just MHO.
> whenever one needs something from the other, they message each other via that file. Tasks can bounce back and forth as issues are resolved and tested.
Why isn't this just one coherent agent loop with subtools/agents as appropriate? If these tasks are related in some way, having a single context would probably make it go much better.
The freewheeling messaging part is where the token bloat is coming from. I suspect that for some of us this is actually the point. I think it's a mostly form of entertainment to do things this way. The next logical step from Factorio gameplay.
Parallel agents remind me a lot about multi core compute. It's incredibly easy to take a single core product and make it run much worse across a lot of cores.
VRAM aint cheap, Sam Altman ruined the cost of memory, Nvidia doesnt make enough consumer GPUs letting the market go insane over them, I still have friends on 1070s or 1070 TIs because GPUs have been severely overpriced for too long. I remember when a gaming PC was only $1000.
Even so why would anyone not sleep on a model they cannot run?
Seriously, if a single politician stepped forward and said "i'll bring down ram prices" they could then shoot a puppy and call me a slur and I'd still go out and doorknock for them.
Look at the financials of these companies, they're all making obscene margins and do they plan to increase production? No. Micron is doing a stock buy back to pump the price of their share.
There's sanctions, tarrifs, and a DOJ who doesn't give a shit. Until we can fix that the insanity will continue. Phones will be unaffordable. Laptops will be obscene. Gaming consoles will be thousands of dollars. Desktops will be dead.
If you're waiting for some David Ricardo equation to happen, tough cookies, it's not coming.
The market is legally locked down and we're in hostage pricing mode.
And what's the story? You can't afford electronics because we're using it to build robots to take your job? I mean ...
Micron has 3 brand new fabs currently under construction, 2 Boise, 1 in New York as the first of 4 planned for a campus.
Plus expanding other existing facilities.
These things take ~3-5 years from breaking ground to full production. You'd have had to anticipate the current demand years before it happened in order to be bringing production on-line before 2030 or so.
Samsung and HK Hynix also have fabs under construction and planned.
CXMT started 11 years ago and only now is reaching any real volume. If they decided a year ago to react to the current demand cycle they'd be 6-7 years out.
Not much you can really do to wish for more fabrication to exist on any timeline not measured in fractional decades.
Could they do more and react quicker? Probably, but everything I've read on the subject seems to point to 3 years is absolute bare minimum if you happen to have a shovel ready project with the land bought, local permitting completed, infrastructure extended to the site, and a skilled workforce already in place. They could suspend buy-backs/dividends today and dump it all into building production and there would be no material impact until around 2030.
> The Micron CEO just recently said this is the exact plan
CEO simply stated the demand pressure will not go away through 2027, and supply will not increase until around 2028 when currently under construction fabs start shipping volume. The article does not support your statement.
Capitalism eats itself this way. Second and third order effects will collapse the demand.
You need to keep the market healthy, not some insane Bitcoin style HODL pump - that's how you get wrecked.
I mean I'm not a neoclassicalist but I've read all of them. I'm in consensus with them here. There's a bunch of theories on what a healthy market is but what we're currently seeing matches none of them.
It's short term profitable but long term disastrous, especially in a world where new mathematics and techniques could literally collapse the demand overnight.
Imagine if some paper hits arxiv and the 256 GB requirement for some model now becomes 64. Woops!
Some clever trick about how attention heads and context Windows work could potentially slash a bunch of requirements by giant margins and all they're doing is firing the starting gun at that global race with every obscenely priced unit they sell.
But if prices were reasonable, this wouldn't be an apocalypse. It'd be fine. Consumers wouldn't rush to 64GB, they'd say " Cool I can multitask now at 256" or " great I can do horizontal scalability' or something else.
But no they created the market conditions so now what would happen is the consumer will immediately flip the 192GB they don't need on eBay, hoping to snatch a profit before the prices tank and the second hand market will be flooded the rug will be pulled out from the luxury pricing and everyone will get screwed.
This has happened in electronics markets before. Many times.
When Engels talked about the grave diggers of capitalism they were looking at it through a 19th century labor/manufacturing lens but arguably this same dynamic is at play here.
What "second- and third-order effects" do you suppose will collapse the demand for RAM? The people complaining most loudly about RAM costs are the people who want to run local models; if that becomes popular it will supercharge RAM demand, because locally-hosted models can't parallelize runs from many users the way cloud-hosted ones can. I don't see any slackening in RAM demand at any point in the foreseeable future, even if the big AI companies all go bust.
This is all hypothetical and debating hypotheticals isn't productive so let's roll back to markets.
Let's say ram used to cost $100 and now that same unit costs $1000. You paid say $500x1,000 for that unit during the price increase or some price where you can currently flip for profit.
You have a very expensive data center and you're in debt financed on the premise that you have these special computers.
Now a new technique comes out and it turns out you only need 1 memory unit for something that used to require 8 or 4 or some meaningful multiplier.
This stuff happens all the time. It's why we don't use BMP files on websites or serve giant MOV files on YouTube. It's why postgres queries are faster now than they were 10 and 20 years ago.
You rent out your machines. You need to service your debt.. Demand may 8x overnight to accommodate but you have a monthly bill to pay and that's unlikely. It's likely going to drop.
Think about it. Your customers are paying maybe $10,000 a month and serving their customers. Now they can drop that to $1,250.
On market if you were to sell some of that ram you have 100% profit right now but not for long.
Jevons paradox assumes unlimited capitalization, zero debt servicing, infinite time horizons...
We live in the real world so what do you do?
Historically the answer has been "sell that shit"
There's an aphorism for this "stairs on the way up elevator on the way down"
If we had a healthy market with sane prices where you can't flip the thing you bought for 100% profit the answer would be "create more value."
> locally-hosted models can't parallelize runs from many users the way cloud-hosted ones can
Why not? Unlike many other workloads, LLM inference actually seems pretty suitable for decentralization (effectively stateless means no availability concerns; bandwidth and latency are relatively forgiving too).
I think locally-hosted models at the org level will definitely be somewhat popular, but you seem to be talking about decentralizing for people's personal, non-business use, and I just don't think that's going to happen to any real degree.
People who say they want local runs really mean it: they want local runs on hardware in their room, not on some decentralized system which, if it existed, would almost certainly just be a worse, less-reliable version of cloud hosting. I'm not saying nobody would use it, but it sounds a lot like things like IPFS, which have also completely failed to displace either cloud storage or buying a bunch of disks for your own private use.
Some people will care a lot about keeping their data on-prem, but many others probably won't, and the former can then resell their spare capacity to the latter.
Decentralized storage is much harder, since there reliability matters a lot more as it's inherently stateful. You have to assume data loss, so you have to replicate everything; with inference, you only have to spend extra resources at failover time. Also storage can't be time-shared in the same way as compute; if it's full, it's full even when not actively accessed.
> The people complaining most loudly about RAM costs are the people who want to run local models
This is a tiny percentage of the population.
Samsung is cutting phone production because of RAM prices.[1] The consumer market is badly affected: budget phones, laptops, general electronics.
The budget segment of sub $100 devices in India has been almost wiped out. Manufacturers cannot afford to spend 50% BOM on RAM+storage. Unless employees are getting a 15-20% wage rise this year, I expect a similar situation in most places.
Between the engineered conflict in the ME triggering O&G price rises, and stratospheric RAM pricing, the situation is pretty bad.
> Imagine if some paper hits arxiv and the 256 GB requirement for some model now becomes 64. Woops!
If you were a DRAM manufacturer, isn't this exactly the kind of thing that would make you think twice about investing years and $billions in new fab construction?
The second Micron boise fab hasn't even broken ground yet, they are still working on the first one. So don't expect these things to be completed in parallel.
Some of my family is pretty happy, though, with the job security as they are pretty convinced these projects are all going to take much longer than what's being stated publicly. Micron is saying the first chip from the new fab will be in 2027... though they also predicted it'd be 2026. The date seems pretty slippy.
Anyone who has been around the semiconductor industry since the last century will remember various huge fabs e.g. in Arizona that were partially built but never finished due to oversupply by the time the walls and roof were done.
If memory prices cool, in about 2 years, Micron will stop new projects, they have done it before.
Especially given CXMT has been able to scale up much faster than what most people expected, only reason their isn't a bigger impact is modern HBM is hard to CXMT even today.
We are likely to see supply double in the next 3 years, but demand even out with optimizations, cooling of data center demand, and most importantly moving some of the dram to flash demand instead which is much easier to produce and scale.
> CXMT started 11 years ago and only now is reaching any real volume. If they decided a year ago to react to the current demand cycle they'd be 6-7 years out.
It's taken them this long to catch up to the DDR5 standard. They've only recently been through qualifications to be a DDR5 supplier for the big boys.
> Every Major Motherboard Maker Now Validates CXMT DDR5
What do you think they'll do? Neither repubs nor dems will touch ai companies in a meaningful way. Anything China does wrt memory fabs week be more significant
I don't have faith in the political parties. Everything is insane. You look at platter recently? It's up 3x in 12 months, not just ssd or nvme, but straight up traditional platter.
After 70 years of decreasing computer prices all of a sudden it's gone 3x, 5x, 10x up in 1 year, we are in total clown world and saying "dur AI" is lazy and doesn't map to reality.
It's Argentina style inflation - as if Honda said "we're only making $500,000 luxury cars now. Everything under $50k we've stopped." and then those cars shoot up to $125k.
It's destroys the market, destroys the consumer, destroys the company, dismantles everything, and they do it for the short term payday.
RAM manufacturers are bidding against NVIDIA and everyone else for the same constrained supply of EUV machines. And it takes years to build more fabs. Micron has multiple fabs coming online in 2027 and 2028.
If we take some time to understand how HBM memory is manufactured (with particular focus on yield risk for final packaging steps), we will hopefully learn that the current capacity crisis is not bullshit.
I guarantee Micron & friends are not intentionally orchestrating their business such that they would suffer a massively reduced chance of yielding on a per-die basis. Unless someone is actually buying HBM devices, they are not going to be making them. These are not a commodity that can be speculatively manufactured in any economically rational way.
>Seriously, if a single politician stepped forward and said "i'll bring down ram prices" they could then shoot a puppy and call me a slur and I'd still go out and doorknock for them.
How many people, outside of tech geeks and megacorps care about RAM prices? And how gullible would you be to BELIEVE the politician they could actually make it happen, and even if they did, that it would extend to the average person, and not JUST megacorps/megadonors?
Phone companies have been differentiating their models based on RAM for a decade. As have laptop and desktop sellers. The reason your router sometimes randomly crashes could very well be a result of not enough memory. The reason it takes such a long time to launch some programs repeatedly is because you don't have enough memory to cache it. Swapped from your browser to an app on your phone, but when you go back to the browser the site has reset and you lost everything you were working on? Not enough memory. Etc.
I think a lot of people care about the downstream effects of memory prices, but I agree with you that they may not realize that they happen because of memory prices.
>The reason your router sometimes randomly crashes could very well be a result of not enough memory. The reason it takes such a long time to launch some programs repeatedly is because you don't have enough memory to cache it. Swapped from your browser to an app on your phone, but when you go back to the browser the site has reset and you lost everything you were working on? Not enough memory. Etc.
That might be true at micro level, but at the macro level more memory just means developers get more lazy with their optimizations, causing apps to get more bloated, eating up any gains in extra memory. There's no reason why slack needs 1+GB to run, yet people are perfectly happy to put up with it.
> Seriously, if a single politician stepped forward and said "i'll bring down ram prices"
Or abolished VAT (the meaning of VAT is that you pay a "rent" for all the infrastructure used to produce the thing) and import taxes (protect your market) on stuff we don't produce in our markets anyway.
The situation is actually much worse and the long term consequences will start materializing soon. The wholesale theft of humanities soul is in progress. It won't be a pretty sight in supposedly civil first world countries, when the human spirit awakens. Currently we are still pressing that snooze button hard and repeatedly, as I think most are keenly and deeply aware of what needs to happen but that too will cost our souls.
I keep arguing that memory needs more competition, and people keep pointing out that it's too slow and expensive to ramp up. But if the threshold to enter that market is so steep, that means it cannot function as a free market and requires regulation.
In this case I think investment in more production is the only option, and it needs to happen even if it is expensive and slow.
He doesn't ruin the cost of memory. Advances in memory size and speed are now in full speed mode. Expect drastic increase in the upcoming years. Big factories are in the making and planned. Gigalab in the US and many others in the east.
Since 2010 we have computers with 16gb as being normal. Finally we are moving into a new era where the standard will be 64gb next year and 128 in 2028. Hopefully we reach 1tb in 2030.
There's no BF16, original full quality weights are quantized already and 510GB.
Then good portion of those weights are n-grams (~200GB) that don't need to be in VRAM.
Then KV cache of that model is super lightweight at ~1GB per 1M tokens. If HBF succeeds, then accelerator with 16GB of VRAM and 1TB HBF/NAND is probably all you need (?).
GLM 5.3 Flash runs fine on two Sparks and Qwen 3.8 Flash Next on one is indeed incredible! I made this 3D game with it in two days using Qwen Code as agent:
Having Flash Next local at 150 t/s with 250k context is a joy. It’s as good as Sonnet 5. It will spaz out but it was less eager compared to DS Flash 4.1. Both are good but I find I prefer Flash Next. This and Qwen 27B are the models people should be freaking out about.
I reimplemented most of the features of the Deepseek v4.1 flash paper (apart from quantization aware training which doesn't make sense because my implementation uses float32 precision anyways)
I'm currently learning how to distill reasoning traces (check my other github repositories) but I think that a locally selfhostable deepseek is possible with my mixture of experts sharding mechanism. I decided to optimize everything for CPU parallelization, with the idea that the KV cache and meta model have to run from CPU RAM anyways, so the experts can also be loaded/unloaded at runtime if needbe, to save more RAM.
My assumption is that the KV cache optimizations in combination with the CED and compressed attention features are the reason why v4.1 flash has so few hallucination problems and such a strong self-lookup/thinking behavior. But that's more a gut feeling, need to evaluate and test this more thoroughly.
Anyways, would love to see someone train this on their own datasets. Currently my pipeline is kinda optimized for parquet and zim files.
The article isn't just about running locally though. The author is saying it's super cheap to run the model through Opencode Go (and presumably OpenRouter etc.) Personally I'm always most excited by models I can actually run locally, but even these huge open source models open up the competitive landscape for companies to let you call models via an API or just lease compute. And they don't have to charge you to offset research, training, huge staffs of the best minds in the world, crazy PR etc. I think that's a big win for customers and buts competitive pressure on the frontier labs as well.
You can run it locally for the price of a decent car, or run it (hopefully) privately on somebody else's hardware at vast.ai or a similar provider for much less. What's not to like?
No, you won't get frontier-level intelligence on a 1070Ti. Yes, it should be illegal to do what Altman did. Since we clearly don't live in the best of all possible worlds, we need to settle, and DS4.1 Flash is a good place to do that.
For tasks that don't require vision I personally like the NVFP4 quant of GLM 5.3 from Local Inference Lab better than DS4.1F, but they are both well beyond awesome.
I did somet math and completely gave up on the idea of trying any worthwhile local model and figured I'd rather pay the 15-30 USD per month via subscription and/or API key combos for years than buying a local setup which might go out of date very fast, if it doesn't goes kaput just out of warranty. I won't be surprised if RAM scarcity is an concerted effort to herd people towards the remote models :)
Allow a question from someone who’s only got a very vague idea of how this kind of stuff works behind the scenes: say I rent usage of this model through one of the many LLM hosting providers out there, and let‘s assume I use it extensively through something like Pi or OpenCode and vibe code away all the time, keeping the hosted model occupied as much as I can, happily burning my credits.
Does that mean that there is a hardware cluster as described by you above that is crunching away just for me?
So at FP16, I alone keep a 1,664 GiB system occupied all the time?
No, a cluster can server multiple users at the same time, providers cap the tok/s so that one cluster can run inference on multiple inputs at the same time. OpenAI with their new ultrafast mode is probably reserving the whole cluster or prioritizing requests of ultrafast users above others with a higher tok/s hence the high price and high speed. There's many other knobs providers tweak that they don't show the users, for example I doubt many providers are hosting the full FP16 version.
The "expensive part" of generating the next token is streaming in the model weights from memory. The computations are relatively simple, which is called a "low arithmetic intensity" in industry jargon.
So what they do is batch multiple chats together and compute the neuron activations for all of them together.
This is vaguely similar to how some database engines work, where if multiple users need to run a "whole table scan" query, the additional users "join" the streaming workload of the first query mid-way, then loop back around to complete the first part that they missed. The AI accelerators don't do this looping, but the concept is the same: amortize the expensive I/O over multiple computations running in parallel.
The "turbo mode" token rate thing is almost certainly your query getting sent to slower or faster hardware, like B200 vs newer B300 kit.
It depends hugely on what "rent usage of this model through one of the many LLM hosting providers" means. If you're asking them to host the model privately then yes, all of that 1.6T of RAM is likely in use holding weights, activations and KV cache by an inference engine that's only getting/answering requests from you alone. When you aren't actively using the model the hosting process is still active and waiting with all of that memory still wired to it.
As background: For the most part VRAM oversubscription/paging/swapping isn't a thing in the same way that RAM for a VM often is. There are some approaches to it, but (to my knowledge) not at that sort of scale.
There are some systemic reasons for this, but very broadly speaking the GPU vendors are building toward the highest bandwidth and lowest latency possible, and the overhead/complexity of something like protected memory modes serves neither of those priorities.
Projects like DwarfStar https://github.com/antirez/ds4 really lower the hardware bar a lot so Deepseek 4.1 flash and other mixture of expert models can run on consumer hardware. There are also other inference providers who make their money serving openweight models. Services like OpenRouter make it all too easy to utilize these models. Access to these models isn't hard. The hardware moat is becoming pretty easy to bridge.
More concretely DwarfStar M5 128GB Deepseek 4.1 flash 1K tokens @ 29s, 5K tokens + reasoning @ 147s, 10k token prompt @ 463 tokens/s = 22s. Hardware buy-in USD$7K / AUD$8.5K / EUR€6.8K. At typical workloads, ROI is still poor vs. current-era subsidies, but owning hardware is good for privacy/longevity/connectivity independence. Whether you actually consider Apple hardware 'owned' is a valid and thought provoking question.
Still gonna take 2-3 years to get DeepSeek V4.1 Flash quality at decent speeds on reasonably priced hardware.
Hardware update cycles are 2-3 years even on the high end, so it's still a ways away before "good enough" and "local" belong in the same sentence for the average person.
And by then, DeepSeek V6 Flash will be too cheap to meter, 5x faster, and 10x better, so... You'd still need to go out of your way.
Most people are spending most of their time on their phones anyway. ..
Flash Next is a basically there. It really depends on what you are doing. This model is great. People forget that they felt Opus 4.6 was a great model and now you have it at home.
"1070s or 1070 TIs because GPUs have been severely overpriced for too long" ... ."
1070ti launch MSRP was $450 ish.
5070 could be had in the last year for 5xx-6xx range easily.
All things considered - (inflation being about 30%~ (guess)) between these two timelines. You are looking at 300% performance difference at a cost dollar for dollar that is cheaper then when they purchased their cards.
Might be a bit of a stretch blaming it on "severely overpriced for too long..."
Are you counting the n-gram/PLE as part of the model weights there? They can go in host memory. Would be good to show your working. Also the released weights are pre-quantised and presumably QATed, so your "Full Precision" and INT8 are simply not a version of the model that actually exists.
Edit: I went and checked for you. The LM backbone is 307.2 GB (286.1 GiB), straight from DeepSeek's upload. The n-gram table is 203.1 GB (189.1 GiB), which goes in host RAM. Note the embeddings are higher precision than the expert tensors, so it's a larger fraction of the bytes than it is of the parameters.
DeepSeek V4.1 Flash is mixed MXFP4/MXFP8 so all but the INT4 calculation is wrong here and that's still wrong because you can run it on < 400GB VRAM. The n-gram table is MXFP8, but can be offloaded to RAM or disk without too much of a performance hit. Really, you could probably cram it onto < 300GB VRAM if you're willing to apply a small quant to certain parts of the model considering how little VRAM is dedicated to kv cache.
I know my comment is a little nit picky because it's still pretty expensive to run, but it's not quite as bad as this comment makes it out to be. Really, if you're VRAM constrained, take a look at GLM 5.3 Flash or Qwen 3.8 Flash Next before you worry about this model as all three models perform pretty similarly.
In nvfp4, it's about 300 gigs once you offload n-grams, 491 without offloading, you can run it pretty well on 4x DGX Sparks, which last I checked was about $20k. So, it's definitely runnable.
Or you can just use any of the neoclouds' shared hosting. The thing for them to be freaked out is that these models are getting good enough very quickly, and all the shared hosting providers can run them for a tiny fraction of what the frontier model companies charge.
I'm paying for the heavily-discounted subscriptions, not the API rates. There isn't really a cost gap for me. DeepSeek doesn't have a subscription to compare to, but when I compared the GLM 5.3 usage I got from a $100/mo Z.ai subscription compared to Opus 5.5 on a $100/mo Claude subscription, there wasn't a big gap. And GLM 5.3 is very clearly not a frontier model (deepseek v4 seemed a lot
closer, but I didn't use it enough to really say for my workloads).
I don't think those subscriptions nave negative contribution margins, either. I think we're seeing a lot of price discrimination by the big labs, and huge margins on their frontier models. The fact that they have been cutting prices to their second-biggest tier of models (Opus/Sol).
Open models catching up and collapsing these margins would worry me if I were a shareholder in the big labs, but as a user, I really doubt that the western labs have bigger environmental impact just because they have higher API costs, I think they have a ton of efficiencies they aren't sharing with customers yet because demand is so high.
The tightening of subscription value has already begun. dsv4.1f is already worth paying for at market API prices. Maybe it goes to 2x because apparently no one has figured out how to match DeepSeek's insane caching efficiency, but I don't see it getting much worse than that.
Plus you can also get dsv4.1f subsidized. OpenCode Go gives 4x if I understand their pricing correctly. Anecdotally, I feel like I get way more out of my $10/mo OpenCode Go sub for the price than my $20/mo ChatGPT, even using gpt-6.1-sol high which is very cheap, and I have yet to convince myself dsv4.1f is a worse model.
It's really not cheaper than frontier subscriptions. It's getting closer, and it's a great model, but it is not more value per task than the frontier subscriptions. Don't be swayed by the token costs, it's very chatty, like 3x more tokens for the same task as sol. I used dsf 4.1 full time for about a week.
It blows frontier API pricing out of the water, but again, look at cost per task, not token usage. Still easily wins though for my work.
I do think it's the most viable alternative I've seen so far, and that applies pressure to the frontier models. Should subscription prices hike or become unavailable for some reason, I know what I'll be using.
When pricing this, it's important to consider whether or not you want to opt out of data training. You won't get the advertised rate. Also the dsf 4.1 subscription providers are throttled af... and of course they are, because otherwise they'd be haemorrhaging money.
DS4.1 flash is $0.30 in / $1.20 out (per M, peak, cache miss)
Opus 5.5 is $4.00 in / $20 out (per M, cache miss)
However, that is API prices.
Anthropic offers a $200/mo subscription. How this translates into usage is admittedly a bit opaque, subject to change, and depends on how exactly you use it. But it's a lot of usage - Semianalysis data shows that $200 is getting you around $2,500 of usage at API rates if you use Opus 5.5. This is close to what I'm seeing anecdotally with my accounts, if anything I have been getting a bit more.
Now, unlike DeepSeek, you can't use your subscriptions to power live AI-driven products, or resell tokens in any way. But for personal coding agents, you can use as many of these subscriptions as you want, for now. So I am paying effectively basically 8% of the published API rates, so at my usage:
DSv4.1: $0.30 in/ $1.20 out
Opus 5.5: $0.32 in / $1.60 out
Obviously, those aren't real prices, but they accurately convey apples to apples what my everyday usage costs me and most other heavy users, and why it's so easy for me to stick with Anthropic/OpenAI.
I don't think it's a coincidence, either - I think the token allowances for these subscriptions are set to be competitive with the open models, so that most coding users (and their incredibly valuable data) stay with the frontier labs, while VC-funded wrapper companies and less-price-sensitive giant companies with strict procurement policies pay exorbitant markups for enterprise contracts at the API rate.
Locked into their tools though. I happen to be very tied to a IDE centric model (old dog can't learn new tricks) and their Desktop thing is a regression for me. I can use the cli, but it wrecks the muscle memory I have with what I use now.
And Anthropic is somewhat unusual in that 10x more tokens via the subsidized path. I imagine more rugpulls are coming.
Not disputing your point in any way, just noting there's already caveats, and more are likely coming.
> It's really not cheaper than frontier subscriptions.
It depends on how you use it. I used to have the $100/mo Claude plan. I would easily blow through limits when I was on the $20/mo plan, but would rarely hit them when on the $100/mo plan.
Lately I've been using GLM 5.3 Flash (from Fireworks), and my spend is $1-$2 per day when I use it for coding, so max $60/mo (less, since I don't use it every day). IIRC DeepSeek 4.1 Flash is priced similarly.
If I had to pay API rates for frontier models, I can't see how $2/day would cut it. Maybe GLM/DS are chattier, but not anywhere near the 10x required to make the price difference not matter.
Sure, if you're running agentic loops all day, 5 days a week, you're probably going to blow past even $200/mo in API charges pretty quickly.
Have you guys see how aggressive is the push for enterprise use by both OpenAI and Anthropic? I had friend from a non-tech industry in Asia telling me that their company was offered free trial of the enterprise version of Claude, with trainings and such.
On the other hand, DS and Z.ai, have zero to none marketing outside China. There is friction to use DS/GlM models and the ZDR is unclear, so most enterprise that has heavy AI usage hasn't move over yet. They would rather spent $200 for the peace of mind than to take the risk of being slam as a national traitor down the road (which again is another form of marketing by Big AI, trying to frame Chinese models as thiefs).
So, I don't think they are not freaking out, it's just that they are addressing different market segments and reacting to the situation differently.
Aggressive marketing (including daily posts on HN). Many people simply don't know about alternatives, there are many people who never heard about pi and opencode and live comfortably in claude-codex bubbles.
I have been using DeepSeek 4.1 flash intensively for over a month. If I run it all day long it costs $1-2. Its fast. Previously I was always quickly running up to my Claude/Codex 5 hour window (on the $20/month plan). The cost savings of DeepSeek is real as shown in this article and I am using subsidized plans.
DeepSeek is horrible at grilling sessions (the /grill* skills to make technical decisions). It doesn't know how to explain things. Maybe the skill could be adjusted. It also doesn't come up with as good solutions as Opus/Sol.
What I use it for is
* the orchestator of my coding workflows
* the tester/verifier of code changes
* the sub agent that explores code or does web searches
* putting together code base research reports
Previously I planned with Opus/Sol/Astra and then I used DeepSeek for coding, and then reviewed with Opus/Sol/Astra. With the cost improvements to Opus/Sol I am trying to use them for coding instead now so there will be less back and forth review needed.
They are all working together in Pi using the extension @tintinweb/pi-subagents where my workflow skill is calling different subagents that use different models.
Luna is cost competitive, but doesn't score as well on intelligence. I do need the intelligence for most of what I use it for, so I am not motivated to use Luna. Haiku also doesn't seem like a competitive price/performance mix.
DeepSeek's own paper advises against using Max, showing that it normally doesn't perform that much better. I am not using it on Max, so that's not a useful benchmark for me. I have seen other benchmarks where Flash does significantly (30%) better than Luna.
It is super bad on a bit more complex workflows and starts repeating same errors with the same tool until the cycle breaker hits.
6 is worse than 5.6 here.
But it is amazing on generating a report on content generated by better agentic models such as DeepSeek or GLM, which both do a mediocre/bad job on reports.
I'm using DSv4.1 in OpenChamber (eg OpenCode) using the Superpowers skills and a lot of custom AGENTS.md instructions to iron out the kinks and I genuinely cannot see a difference between it and Opus and I've been building native iOS and AppleTV apps, Go servers, Typescript, Cloudflare workers, Svelte/Astro, etc.
It's a super capable model all around from my experience.
How are you getting down to $1-$2 per day running "all day long"? I've been using GLM 5.3 Flash and I also spend $1-$2 per day, but my use is pretty modest, I think. DS 4.1 Flash is priced similarly to GLM 5.3 Flash; can't imagine DS is significantly more token-efficient.
While I agree, this particular discussion chain really frustrates me.
1. "This Flash model is really smart. Here is an article to discuss how smart it is. Why aren't people freaking out about how smart this Flash model is?"
2. "I tried using it for a smart thing. It doesn't work so well for it."
3. "You should know better than to use Flash for smart things. It's not meant for smart things."
Because DeepSeek is not "a month or two" behind as claimed in the article.
These open models still did not beat February's Mythos / Fable 5.
DeepSeek 4.1 Flash is behind GPT 5.6 Sol, and that one is left in the dust by the excellent Opus 5.5.
Rumors say Anthropic is holding in reserve the big improvement, Fable 5.5, for the IPO.
It's plausible that open models are 6 - 12 months behind, and there is no "good enough". As long as progress doesn't slow down, leading labs have nothing to fear.
I am coding CRUD apps with a mix of astra, sol 6.1, fable and opus 5.5. A more capable model would still benefit me imo. Being able to follow high level guidance better, and being able to harness other models for each task would be a big improvement.
Do you know how what you're doing, or do you find yourself working on things you dont understand and need the best model because it's the only way to push your own capabilities (because you're avoiding learning how to do the thing yourself)?
Not asking to be mean, I just genuinely dont know why you'd need the frontier for basic applications.
It's a matter of bandwidth. The more I can offload onto the model, the more I can accomplish. For example, I had to do a lot of security work over the last 2 weeks to get ready for an event. This requires handholding current models on many fronts, like:
1) Do they actually implement the security fixes correctly.
2) Do their fixes create any new edge cases.
3) Do their fixes compromise existing interfaces or API surfaces.
I cannot trust current models to find all the necessary context, or to make what I consider to be good trade offs. A much more capable model would be able to see my existing patterns (or at least not have context rot make them blind to my convention docs) and make trade offs I agree with much more consistently, and I'd be able to do more with my time.
I've actually found models to be pretty poor at driving things I don't know well, so I generally don't do that unless its general design/product exploration and the end product code is throw-away.
Many people would, and you'll find that they're building crappy webapps where you dont need SoTA. Like seriously who needs these frontier models?
Unless you're doing some extermely difficult post-grad lvl research, you do not need a 100x PhD research assistant, especially not for whatever silly SaaS product most people are building.
There's people at my job that get so much more done than everyone else using Fable/Opus/Astra. and all they use is the fastest cheapest models. I'd say the people who are using sota models for everything are doing it just because they prefer to be lazy.
You simply do not need these frontier models, they outgrew most people's needs 6 months ago, but for some reason people still want to run a 700k rack of gpus full throttle to center a div for them.
I agree that for average web dev tasks the open models are already good enough. I've had good experiences with both DeepSeek and GLM. And these models are just better for anything security related since they don't throw massive hissy fits.
However I do actually have a project where I need the frontier models--I'm working on a deep learning project of moderate complexity (something novel/state of the art within its domain, adapting a known approach from published research in a related domain). The difference from Opus 5 -> Opus 5.5 was huge for my project. Opus 5 was struggling, Opus 5.5 is doing really well.
I think the demand for frontier models will continue to be there, at least for a subset of tasks, although I agree that it is probably going to shrink as the non-frontier becomes more and more capable.
I think the market would be huge, especially if it's the "can complete a large task in 3 turns instead of 15" kind of smart. Lots of people and companies would pay for quality + speed.
I think there would be a market, but it would mostly be a FOMO market. That is, people would be doing tasks on it that the "regular" model is more than capable of handling, because they're afraid they're leaving something on the table by not using the absolute best option.
Certainly, there's a real market for it too, with people who would actually use its advanced capabilities, and see the 100x price as worth it.
But sure, even a mostly-FOMO market is still a market. If people are paying, people are paying.
What does 10x more capable look like now? Surely at some point we will reach an asymptote of what can be done purely digitally: all useful coding tasks can be automated, most math research, etc. At some point the physical world becomes the dyke holding back the singularity; until these genius models can scale their investigation into physical experiments and manufacturing, the future will have arrived only in the digital world.
I think you're making a mistake in thinking the digital world is the only one reachable to AI. Robotics and sensing would be opened up by a sufficiently capable AI.
I agree: large market. I think the future of these frontier labs is selling exceptionally powerful and exceptionally expensive models. They'll be used for precision, high value tasks. The rest of us will be happy with good enough and cheap models.
There absolutely is "good enough" and I agree with this author: DeepSeek 4.1 Flash is plenty good enough for all the things I would trust an AI to do at my job.
Agree, will see after the dilution and CoT hack fixed, will they keep the pace now. MiMo had some good numbers recently because it's discovered that the post evaluation RL directly exposes answers to models, so RL and evaluation is runied.
Perhaps on certain benchmarks and for certain work, but anecdotally I've not been able to see a difference between it and Opus on a lot of dev work (web, Go, iOS/AppleTV native, scripting, general tasks)
Imo deepseek 4.1 (and a lot of the cheaper models, Luna is similar) show the issues in benchmarks. At this point.
In actual day to day development the differences are a lot harder to spot. Maybe deepseek is worse, but I asked it to run until it was able to launch itself and verify it worked as expected, and it did. Maybe it wasted some turns, idk, but when it said it was done, it was done.
I have no doubt there's things it's worse at, but what percentage of development is truly novel?
Personally I find this shocking. I don't think our application is that complicated, (typescript full stack graphql reactnative etc) but deepseek 4.1 flash is a bumbling fool, junior-level at best, who takes a very long time to make a very big mess. Opus 5.5 one shots truly impressive code in 5 minutes, while deepseek 4.1 flash takes 20 minutes to do horribly. I simply don't understand how folks claim they get good engineering out of it. Maybe we still care enough about the fundamentals to notice the mess ...
Are you using OpenRouter? I’m honestly surprised open model labs haven’t been calling them out, but heaps of providers either silently serve heavily quant versions, or don’t have inference set up correctly and don’t run the model properly.
Was a night and day difference going directly to deepseek api
The big labs' financials are based on their products being used widely by a lot of the general public. If it turns out that they're actually selling a premium product to premium-product consumers at a premium price point (while everyone else buys DeepSeek-like cheaper/worse products), that's a big issue for them.
If a consumer computer hardware company launched by promising investors that it'd be the next Dell/HP and it turned out to be the next Apple (talking Macs here, not phones or apps/services), that'd be an issue for them too.
> These open models still did not beat February's Mythos / Fable 5.
On what task? By who? On what benchmark? How do you measure in you own workflow the “betterness” or “more goodness” of these or any models? If you don’t say those things you’re just writing a bad ad copy.
> It's plausible that open models are 6 - 12 months behind, and there is no "good enough".
Anecdotally, a lot of people - including myself - seem to really notice much difference between the model now or six months ago. So there really seems to be good enough. It depends on the task you use them for and how you measure the output. For most tasks you really do not need frontier capability. Also how do we know how much of these “big improvements” come from the harness and tooling rather than the raw capability of the model?
Did you see any person claiming an open model was more intelligent than Fable 5?
It's always some sort of "I don't notice the difference".
And honestly, if you don't see a difference between the SOTA from 6 months ago, which would be GPT 5.4, and today's Opus 5.5, you would have to be downright blind. Not sure what else to say - the results are obviously different for any kind of meaningful output.
> Also how do we know how much of these “big improvements” come from the harness and tooling rather than the raw capability of the model?
By simply running the old models in the latest harness. Which none of the people who argue "it's all the harness" ever do.
The benchmarks don’t mean shit. Opus 5 was a terrible model and yet it had very impressive benchmarks. All the labs are benchmaxxed to the tits, only open models’ benchmarks are even worth paying attention to because they literally cannot cheat.
Oh man. v4.1-flash has been an sbolute game changer for us. We run all our Personal Assistants now on flash (thinking high) by default and it works incredibly well. There is really no need for basic agentic tasks that might require Kimi K.3 or GLM-5.3 levels.
Once its gets juicier, we let flash launch specialized subagents with specific models. GLM-5.3 for coding or Kimi K.3 for research and critique.
But as a main driver. I love flash. And it brought our bill down by A LOT :D
> Today's models are now good enough for high-quality unattended tasks. Chasing the latest and greatest is silly. It is fun to see the new Fable capabilities, but the tasks we throw at them are usually ridiculous (maybe even insulting) if you believe in LLM sentience. It's like asking a math PhD to organize the files on your desktop.
I'm using DS V4.1 Flash as my main model since their release and it works great for all my coding tasks. My setup is OpenCode Go subscription and obra/superpowers skill.
The only times I try to change models are on general planning tasks (like research this codebase for tech debt mitigation opportunities) or if I need deep research which would benefit from searching the web, in which I still think Gemini is still the best because of the speed and access to google search index. But these are not even 20% of my daily tasks.
Just use a provider hosting it in your country especially if your country has major data centers then its the same as using Anthropic or GPT of GCP Model Garden or AWS Bedrock
nobody here is talking about running frontier level intelligence locally so if you’re Chinaphobic and prefer layers of corporations siphoning your data in between you and the party there are plenty of options instead of directly to the party
Not shilling for them but Ollama cloud hosts domestically with ZDR afaik. I run 95% of my open weight inference through them. The rest goes through Opencode Go $10 plan (which is enough to run 3 hermes agents on DSF 4.1 and leave plenty of left to experiment with when new models drop).
Cloudflare doesn't retain request bodies by default, and if you don't trust that then you shouldn't trust the third party AI provider either. Cloudflare does cache responses, but that doesn't typically apply on API endpoints and is trivially disableable.
What is the cost of access like for DeepSeek-v4.1-flash, compared to GLM-5.3-flash via ZAI's Coding Plan? Because that's what I use; and often hit the "wait". I wouldn't mind trying a new model subscription or even API access which hits around glm-5.3-flash level weight class (which seem to be enough for me; with quite some human suprvision and nudging) but gives muuuuuuch moooore tokens for the same price.
I have a pretty large, complex project I've been building with heavy AI use (new language + compiler). I was following a 'strong model as orchestrator launching cheap models as implementers' pattern, but I recently trialled just using Deepseek-V4.1-Flash as the model for both layers because of the cost savings (with mimo v2.6 flash on code review agents for some decorrelation).
I was previously using GLM-5.3 as the orchestrator, after switching to DS anecdotally there was an unnacceptable quality loss, mostly around not taking all the relevant context into account when making decisions, pulling new design out of thin air without discussion too often, and being way too wordy and rambly in documentation despite prompting to avoid it. There's a lot of docs, rulings, core concepts, design philosophy to uphold and DS was just not cutting it.
However, it's perfectly capable of being the sole agent for all of my well specced implementation tasks. I've gone back to GLM as the orchestrator.
On the Claude side of things I was previously following "strong model directs weak" with Fable directing Opus/Sonnet (its choice per-task). Since Opus 5.5 came out I've just been having Opus direct Opus.
The sub-agent separation is still valuable to keep context clean for the orchestrator, but I just have no reason to use Sonnet as the grunt-work implementer because I'm finding it hard to run out of tokens with Opus 5.5 on a $200 subscription plan. It's really really good at subjective quality of work per token used.
I would probably go that route if I could use other harnesses with claude models, but I don't want to be locked in to claude code, and their API pricing (which you need to use it with other harnesses) is so much higher than subscription.
I have actually just dropped to using sonnet for everything, sure it does need some directing but I have yet to see a need to jump to opus.
To me it feels like sonnet/terra and composer 2.5 and grok 4.7 are actually good enough for most tasks and these companies are pushing the high models simply to make money.
You can just use plenty of good handmade languages now, I'm not claiming mine will ever be good or useful to anyone other than myself. I'm using AI so heavily on my project because I want to explore the PL design space without spending years on implementation for things I'll probably want to throw away and rewrite a different way once I actually play around with them properly. And because I lack the motivation to persist through the sheer volume of grunt work that language implementation needs to get to the juicy interesting parts.
I used Superpowers for awhile, but I don't feel like it gave me significantly better results than just raw-dogging it. It did however burn through my tokens significantly faster.
Maybe I'll come to miss it now that I removed it, but I certainly don't yet.
Its simple. My company is very willing and able to pay ~$200/engineer/month for the best version of these tools. My company is even willing and able to pay as much as $500/engineer/month, but does not need to at this moment.
My company is not willing to pay $50/engineer/month for a cheaper version that is nearly as good. My company is also not willing to pay any amount for a product produced by China, even if it is hosted in the United States.
It’s still cheaper to get a subscription to an agent harness with frontier model backing for most people. DeepSeek really only becomes appealing to me when I want to do something the monthly subscription harnesses don’t support (API access) or more harshly charge quota for (like running coordinated sandboxed subagents). DeepSeek is then great because of the low price and price transparency, it just can’t compete with subsidized monthly access.
I actually have a fairly simple answer to that: if it doesn't come up in the list of LLMs that Cursor supports, it effectively doesn't exist.
I'm well aware that there's nearly infinite opportunities to yak shave "perfect" OpenRouter setups and some people appear to enjoy bouncing from IDE to IDE as though change costs aren't a thing, but I discovered that I genuinely like Cursor and at least right now it's insanely subsidized by Auto clearly defaulting to whatever Grok's most powerful model is.
I dropped my $200/month subscription to $20/month and stick to Auto for all but really important Plan tasks, and I have basically zero chance of using up my monthly credits even using it 6-10 hours some days.
Because most of the people use it through enterprise agreements and don't pay the bill?
I run it for my own use cases and its pricing plus caching capabilities are hard to beat, cents for millions of tokens.
https://substack.com/@rubenafo/note/c-332218129?r=26y5kn&utm...
The problem with this low listed price per token is that, in reality, DeepSeek 4.1 Flash uses 10× more tokens than GPT-6.1 Sol for an equivalent task and delivers a lower-quality result. So there’s no real benefit to paying 10× less per token. Also, as someone else pointed out, OpenAI and Anthropic currently offer subsidized subscriptions for $100 or $200 a month that provide far more tokens than the API, so we should take advantage of them while that lasts.
That's my experience too. sol-6.1 goes straight to solution, like it has done it 100 times before. Deepseek will explore and insecurely overthink like it's an intern made CEO.
You can see this in the Artificial Analysis benchmarks - GPT 6.1 Sol on medium thinking scores 10 points higher than DeepSeek 4.1, but is actually cheaper per task, as it only outputs 15 million tokens instead of 250 million.
Though for longer sessions I think DS4.1 would still come out cheaper... it's hard to beat that 98% cache discount
I had the same experience using ds 4.1 last couple of weeks. It’s insanely good for the price. I’m doing mostly web dev it excels at everything I throw at it. The pricing is ridiculous. I canceled my gpt subscription and haven’t looked back hope the pricing stays like that. I almost never need a better model. I still keep my Claude 20$ sub for now but I feel like one more iteration and I won’t need even that anymore I hardly use it
If DS4.1 impresses you I would be really interested to see your comparison to GLM 5.3. I switched from the one to the other and even if GLM 5.3 is a bit slower I don't think I'll be going back.
DS4 (not 4.1) crossed my dont-care threshold and I genuinely stopped paying attention to new models. I'd love to try GLM 5.3 but I just don't see any point in spending the effort any more. I can get passable intelligence for a bargain price either direct from China or from a ZDR EU provider for a small markup. Paying 10x more will not make me 10x happier, it's unlikely to make me even 1.1x happier now I've got some intuition for the natural limits of these models.
I don't even bother checking how much I spent on API any more, its well under $30 over the past 2 months despite daily constant use. Who even needs a subscription at these numbers?
Wallclock time matters and GLM 5.3, even when it is considerably slower (~30 tokens right now vs 100+ on DeepSeek 4) it is quite frequently faster on the same set of tasks overall. Deepseek 4 seems to do the 'Oh, wait' thing just about forever and has a tendency to find irrelevant rabbit holes that it then spends a massive amount of tokens on.
The question seems rhetorical but I think there are two reasons in some combination. First is it there is some awareness lag here. That lag can be on the producer and consumer side. Software enterprises are pretty slow to adopt new things and slow to try new things so they might only be aware of openai and Claude as options. Plus there are some scariness because deep seek is a Chinese model and therefore export restricted - never mind that there are American in European providers.
The other reason is more interesting. Maybe the frontier providers think that price performance is irrelevant in light of very powerful frontier models that can start the RSI loop and or a huge displacement of work and a winner take all economic situation. After all if frontier providers earn everyone's money then you won't have any money to spend on any model 100x cheaper or not.
I think it also has a bit to do with the AI sector of tech still moving at lightning speed.
Theres already models that outdo DS 4.1 flash in cost/performance. Luna 6 on max effort for example. Luna also doesn't care what time of the day it is for cost calculation.
And I'm sure by the time people ask why Luna 6 is being slept on there will be another cost/performance king
Luna is very slow and bad at agentic tasks. DS runs circles around it and there are US providers providing cheaper rates no matter the time of the day.
While using DeepSeek v4.1 Flash I was architecting a system and I made a mistake of drawing the RPC boundaries at a wrong place that did cost me in so many ways.
I realized that mistake and guided DeepSeek where it should be.
Next I fired Fabble 5.5 set to high to check if the hype is real about Fabble. It exhausted 89% of quota and came up with NOTHING that DeepSeek hadn't flagged itself already in its notes.
Do you mean Fable 5.1? Or Opus 5.5? I'm not sure what you're working on but for me DS 4.1 flash isn't nearly at their level. For the price it's obvious very impressive, though Luna 6.0 is excellent too.
The problem with benchmarks and proprietary models is that one day a model is best at doing X, another day that's not so sure. And anyway, we are not throwing the same X.
I've found supposedly smaller and, less performant models do better on certain tasks. I end up using several models, sticking to what my unconscious statistical observations tell me to use for the kind of task at hand.
I am 4.1 maxxing on commandcode GOAT Plan + api rates with oh my pi for the last 4 weeks, it's absolutely amazing and crazy fast, it's alright if it makes a mistake, I have enough time to iterate again, I have also added an advisor layer of mimo 2.6 pro which does make it a notch smarter. Getting haiku 5.5/sonnet5.5 to work on plans and letting 4.1 flash work through it is helping a ton too.
I am a big ChatGPT fan, all our team has ChatGPT Subs, but the TPS across all models including luna is just so damn slow.
Commandcode giving 60$ worth of Deepseek for 10$ is just genuinely goat.
And it never says no for cyber tasks so that's a big win
Comparing it with Anthropic, ANY model is cheaper and more effective. Don't get me wrong, Anthropic models are good, but they're always more expensive for the same task, even compared to other closed source models. At this point, I honestly think Anthropic has played the nasty trick to fine tune the models to be too verbose and charge us for more tokens.
Not sure why the author thinks Anthropic's models spend more power and water than DeepSeek, there's no evidence of that. Their pricing has more to do with premium perception and less to do with COGS.
Everyone does optimization of model serving because it's good for every player in there.
(Also the water consumption thing is not a real issue.)
Because it’s not even that cheap? The author chose to only include Claude in their chart and ignored the fact that 6.1-sol and even more so luna can easily beat Deepseek on cost. Of course almost free cache used to be the main differentiator, raw token cost is deceptive since 4.1 just uses way more tokens than most other models
> Sure, they stole Claude's training, and Anthropic stole it from other people. I'm not getting into the whole who-owns-whose-data debate, because most developers aren't thinking like that. They're just trying to get the most bang for their buck.
Until they get laid off and suddenly discover their moral compass.
> With my OpenCode Go sub of $10/month, DeepSeek is basically unlimited
My OpenCode Go monthly window was scheduled to reset this morning. It was sitting at 22% used despite me using DeepSeek V4.1 Flash heavily as my implementation agent the past couple weeks (I use gpt-6.1-sol high for planning/orchestration).
I had 1.5 hours left so I fired up first 10, then 20, and finally 50 concurrent subagents all working on reverse engineering C code from an old PC game. They found over 100 new functions.
This is the first workload I've found that could make a dent in my sub. It got my 5 hour window to 85% used, but sadly my monthly was still only at about 35% when it reset. So that cost maybe $2.
How are you guys doing orchestration? I have been fumbling around in my free time trying to build something for myself but is there a repo or something that just works?
I might not be the best person to ask. I use Pi harness in tmux and just ask my current agent to spawn interactive pi instances in new tmux windows, create a sentinel file for each of them, and monitor the sentinel files for signals every 2 seconds.
Currently have auto compaction turned off. When the orchestrator's context is getting close to full, I have it write a handoff markdown file and point a fresh agent at it.
I do feel like I'm getting close to the point where I might be ready for something more sophisticated, especially wrt to subagents communicating with the orchestrator.
That limitation is what stops me from using Pi for anything serious. I may have to configure it and then configure it and it will eventually become a codex, a claude code or so. I recently heard the maintainers added mcp to it (in stock, not via plugin), I wonder what stopped them from adding subagent function, and decent loop capacity to it.
> Currently have auto compaction turned off. When the orchestrator's context is getting close to full, I have it write a handoff markdown file and point a fresh agent at it.
In a good harness that should be how auto compaction works anyway
I think because we're all just using it thinking we have found the "model for me" and never mentioning it to anyone because what would we say? It's good. It's a bit like the Logitech MX Master, as more and more people assumed they had found the ideal mouse for their purposes, it quietly became the professional standard through sheer adoption.
1. a model that works for one person/task may not work for another;
2. there are many models (DeepSeek, Qwen, GPT, Claude, Gemini, etc.) that are released every 6 months or so;
3. it takes time to use, test, and evaluate the suitability of a new model and not everyone has an automated evaluation process for their use cases.
Thus, if you find a model that works for you then you are not going to spend more time evaluating a model that may not work, or may only do so when time permits.
Personally I have not used anything but 4.1 since it came out. I have a dataset that turns any model into pure hallucination machine, not only DeepSeek does not hallucinate, it builds new insights by combining its insights. It's not only cheap, its far better (at least for me)
No doubt about it, that's why their push for international regulation to the levels of nuclear inspections using the narrative of annihilation and apocalypse
I think the interesting provider to cross-check this assumption with here is Meta, who is clearly freaking out, and is currently providing Muse 1.3 even cheaper so long as you are willing to share data with them
Technically yes, but has been reported to be quite benchmaxxed. In practice Deepseek Flash 4.1 and GLM 5.3 might therefore still outperform Mimo 2.6 pro.
I’ve been using it a lot and it’s performing really well. Not GLM 5.3 levels but it beats Deepseek for my use. I’ve used it on long running coding tasks, though mostly prototyping, but it’s done a great job at very low cost.
I did a test on this a couple weeks ago. What I found was that the chinese models were far better than API rates, but about comparable on price vs the subsidized subcription model (chatgpt). Also it was my experience that codex completed tasks quicker.
That said, it's my best understanding that these american companies aren't profitable and will eventually raise rates (the old uber trick) so I'm keeping myself ready to switch when that day comes.
I keep a spreadsheet that estimates actual value (dollar amount per token per month, per subscription rate limit) and open weights are basically always cheaper than frontier weights. Recently things like GPT 5.6 Luna finally got the frontier close to the value of open weights but their limits keep them behind.
Enterprise is not freaking out because DeepSeek 4.1 Flash does not actually occupy a spot on the Pareto frontier for non-coding enterprise workflows. We see this at my employer, focused on non-technical knowledge work. Luna 6 and now Haiku 5.5 are both very competitive if not better on all axes that we care about
I've been focusing on deep research related tasks for biotch and life science applications. The problem with this sort of task is that we need subagents to reason through multiple (potentially 100s or more) chains of knowledge/concept/evidence, so the token usage really explodes as complexity of the task and data expands. A typical task can cost me nearly a $1k overnight...
I've been testing out GLM5.3, but now I'm really tempted to try to Deepseek 4.1 flash too. Any chance you've benchmarked / compared the two?
Been using Flash 4.1 via the ante harness to blast through a GBA recomp. The ante team has pushed hard to make Flash 4.1 perform well under it. So far, I've maybe spent $10 over the last 3 days. Its a real workhorse and works much better in this harness
I didn't know it was so good. That was a gut punch. But I'm pretty sure the market already priced this in. And people are rightfully concerned about the owner of the data. Anything that is concerned with social, political or financial data goes out of your country to another one and is maybe even kept as a potential weapon.
It's quite good and the labs are definitely scared, that's why they are lowering API prices and continuing to subsidize subscription plans aggressively to keep anyone from using this stuff.
That said, has anyone else found DS models to be unpolished? They seem to "lose their mind" a lot more often than Claude/GPT. I have tried all of the top open source models that came out over the past ~4 months or so and the GLM models (5.2, 5.3, and 5.3-Flash) have been much more usable for me. They feel like Opus but X months ago, DS feels like something else.
I've been on the API-only mentality for months and all the open models were definitely the stuff I loved the most. Kimi K2.5, 2.7 and Deepseek v4 were among my favorites, while sparingly using Opus or whatever OpenAI had for specific situations.
But ever since I've switched to one of the $100-tier subs, I can see why a lot of the people on it don't really discuss the open models often. I'd still use it especially when it comes to sensitive inputs, but for most work, what you get on OpenAI or Anthropic is really more than enough.
It really got even better when they also made their cheaper models up to par if not better than the open models.
I do think the crowd for open models are out there, especially when you see trillions of tokens running for them on OpenCode or OpenRouter leaderboards.
One could argue that all this whole "Pacing the Frontier" bullshit is the industry freaking out regarding the danger of open models.
P.S. That's not to mean there aren't dangers regarding AI. I just don't trust the people making money from selling AI to manage those risks ethically instead of "protecting" us from those risks like pimps but with suits and good manners.
I’m also a heavy DS 4.1 Flash user—especially when it’s available at those off‑peak prices, which is an awesome deal. And, like you said, it’s genuinely powerful and very snappy. I’m planning to evaluate the differences between `reasoning_effort` settings today.
Not sure "freaking out" is the word I would use, but it’s fairly obvious looking at OpenRouter usage that the price cuts on Luna a while back were in response to intense competition from dsv4.
So the industry is responding, where it matters. Which is on heavy API usage, not coding subs.
Luna and Haiku 5.5 are just as cheap and much better.
Don't really understand people who say DS4 or 4.1 have frontier level performance. Anyone who has used it will tell you that it's a hallucination factory. The only thing it has going for it is deepseek's unique infrastructure that allows better cache retention, but the cost savings from that obviously come nowhere near how much subsidized usage you get out of even a $20 subscription with openai or anthropic.
Both Luna and Haiku are more expensive than DS4.1 Flash at API cost. But ignore that; nobody should be using API. The open weight subscriptions for DS4.1 Flash provide higher limits at lower prices than OpenAI or Anthropic subscriptions. Finally, Haiku actually is far less token efficient/ outputs more per task. No matter how you slice it, the open weight is cheaper.
Also consider that for things like cyber work, the frontier models give you nerfed results and poor performance. Whereas the open weight isn't nerfed, and I routinely get 200t/s with my subscription. Finally, DS4.1 Flash is natively multimodal, while Haiku isn't.
They're all perfectly fine models, you should use any one of them you want. But DS4.1 Flash can do more for less. (That said: GLM-5.3-Flash is even better and cheaper...)
There is no subscription for 1st party deepseek models, the ones that exist are all fronting as middlemen for discounted openrouter providers that serve quantized models with worse cache retention. The only real option for deepseek has been API for a few months now.
oai and anthropic also subsidize the hell out of their subs compared to what you will find in smaller competitors, a $20 codex sub gets you like $100-150 usage/wk which goes way further than 2x opencode go (which would only net out to $120 of deepseek 4.1 usage a month, on top of being low performing quantized trash).
Yeah. $100 for Claude just about gives me all the usage I want, as a more or less full-time hobbyist having it work in the background most of the day. I was trying to economize by having a local LLM, then Deepseek, then Cursor/Grok, and then I got a taste of Opus 5.5 and I simply cannot go back to having to carefully spec things out and double-check work. I just let it decide, Opus or Sonnet for the next task, and I get almost perfect results. Probably similar with OpenAI's models.
The token-equivalent monthly spend is > $5K+. If Deepseek's token cost is 20x cheaper, that's $250/mo, and I'd be spending a lot more of my brainpower babysitting it and getting worse results.
For business/team accounts that pay per-token, maybe I can see the "freaking out" being warranted on the part of the fronter labs. But as long as they're willing to subsidize their end-user subscriptions, I'm not going to move off of them until the alternatives are truly at their level.
That's what I thought until August. Used it for half a year almost exclusively. But after the API price increase I'm back at the Claude Pro and Kimi subscriptions.
Not only the model but the hardware it is running on. The Huawei chips they are using instead of NVIDIA are vastly less expensive. China is going to scale past the west. The idea that they are “only a few months behind” is today and many of us cannot even bring ourselves to admit it. The future is even more dramatic.
I guess what we're seeing is selection bias - people clicking on this HN story will be those who are interested in DeepSeek. And those people who are invested in DeepSeek may not like the facts that you presented.
It's annoying that social networks work this way. The upvote should be for high-quality content and the downvote should be for low-quality content. But .. well.. human nature and tribal dynamics always seem to win.
I have no idea if anthropic can actually make money at their subsidized subscription rates (you can easily hit your monthly cost in one 5 hour session if you price out the tokens through the api), but if subscriptions didn't exist, I do think everyone would be on deepseek 4.1 and just not look back.
DeepSeek 4.1 Flash 0910 is perfect for M5 Ultra 256GiB. Running it fully resident in RAM, prefill at ~2500 tok/s and decode at ~40 tok/s. Probably tons of room to improve from there.
Because it hallucinates a lot. There's no free lunch. Although the new architecture is a genuine move forward. The DeepSeek guys are really top notch researchers and devs.
I just keep getting more ambitious with what I use AI for; and that type of work needs surfing the frontier at all times. Because ultimately, many capabilities are not yet saturated
Maybe this is a minor issue, but it seems like different providers on OpenRouter etc. have different quant settings. I imagine that that affects the perception of the model quite a bit.
they should be freaking out because every time the chinese labs or non "frontier" labs release a model that is only a few months behind and much cheaper than the openai/anthropic models it shows that they don't deserve their valuations
Perhaps, OR it might be that most people in the markets (think that they) are not all that exposed to the valuation AI labs and so their eventual collapse doesn't matter.
Or perhaps they consider the upside from cheap Chinese models to hedge the effect that OpenAI/Anthropic collapsing would have on their portfolios. This would make sense for (hedge funds holding) most companies: they don't really care about who supplies the AI, as long as they get it at roughly the same price as their competitors.
ironically, at my large enterprise, they aren't yet distinguishing between "chinese models" and "chinese models hosted at microsoft foundry". so far it's just _banned_. i'm not at all pretending it's like that at other orgs.
People tend to conflate the question "is AI a useful technology?" with "are the AI companies going to do well?" but they're surprisingly separated in practice, with either one able to be true while the other is false. There is a lot of money tied up in a lot of hardware with a lot of loans made against that hardware as collateral all based on the assumption that AIs are going to need more and more and more and more hardware and whoever has the hardware wins. If a much better model comes out that requires vastly less hardware, or even more accurately, merely charges vastly less than the current AI companies, then to a first approximation (barring Jevon's paradox, and bearing in mind there's no timeline guarantee on that) all that hardware becomes much less valuable for being grotesquely oversupplied relative to what is necessary, and even though that would generally make AI objectively more useful than it was before, it would cause mass financial chaos in the markets.
The markets need a very particular rate of progress. It isn't entirely clear to me that it's even a possible rate of progress, it may be overconstrained, but they certainly don't have plans for the AI models to get commoditized on the timeframes of these vast, vast array of loans being made against hardware as collateral. Spend a metric shit ton of money to kill all your competition then charge monopoly rent on the one thing absolutely everyone needs doesn't work if you can't economically "kill all your competition" because the economics favor them in the spending spree.
And then, based on the fact that this is not even remotely complicated logic, there are plenty of people who are fully aware that they have a lot of money tied up in not running around telling everyone how wonderful the cheap models have become.
It's also unclear whether those who approved those loans understand GPUs depreciation. In any case, progress in software but also hardware could bring chaos and ruin their house of cards.
> It's also unclear whether those who approved those loans understand GPUs depreciation.
This also assumes heavy utilization, though. If there's heavy utilization, it might mean they're doing well. If they're all spinning, it's time to raise prices.
Agree that people leaving the big two companies is going to be hard to keep a pulse on prior to IPO.
Anthropic and OpenAi are in the news, so they get the press and people go and try out their product. Large enterprise businesses are going to make larger, longer-term contracts with them and are only going to pivot if they think switching costs are easy or if they think the provider won't deliver.
The other inference producers are less well known or you need to get your cloud sales rep to tell you how to switch to them as a provider rather than Anthropic or OpenAI.
I use OpenRouter, I know switching is easy, but larger businesses tend to work in yearly cycles. DeepSeek v4 Flash came out in late April.
I agree OpenAI and Anthropic are going to struggle when the median price of running a smart-enough model keeps falling.
Edit: I also think demand for hardware will be rapidly absorbed by other companies if Anthropic or OpenAI stumble. We've finally turned hardware directly into runnable intelligence and people are not going to go back to the old ways.
> And to the self-hosters out there, the economics of 4.1 Flash mean self-hosting is not worth it. If saving money is your goal, you will never recoup the costs.
Self-hosting is, for most enterprises, absolutely not about economics but rather about data confidentiality.
And in that regard, yes, the open-source weight models, especially the chinese ones will eat the fat closed US model's lunch big time.
I've been using DeepSeek's API and have been happy with it, but I might look into OpenCode as well. Does OpenCode run a quantised version or use different providers from the official one?
It's a solid little model, and I appreciate DeepSeek's commitment to the bit in releasing a brand new pretrain, double the size, numerous architectural innovations as a ".1" release over the excellent DeepSeek V4 Flash.
That is opposite to my experience so far, can you describe your coding tasks? Mine are systems level code, utilities, operating system code, networking and real time control stuff.
I also prefer GLM-5.3-flash to DS-4.1-flash, but it's close. Since Z.ai has been offering essentially free GLM-5.3-flash tokens on their coding plan between 8am-6pm pdt I've been using it a lot... though that ends on Oct 10 IIRC.
Opencode Go gives only 6300 requests for glm and 23 000 for deepseek. And, if I wanted to, I would be able to do all my work on $10 plan with deepseek. It’s very cheap.
Unfortunately OpenCode Go is garbage now. Charm Hyper provides deepseek-4.1-flash at $0.33/$1.31/$0.03/M/cache, and glm-5.3-flash at $0.16/$0.54/$0.03. The rate limit is the same for all models, just the cost; not like the absurd multiple-levels-of-price-rate-limit OpenCode Go pricing, nor their terrible performance.
Deepseek V4.1 Flash is a hidden gem, really. Not to mention, you can easily get it through many providers that offer zero data retention and consistent speeds above 200 tokens per second!
I built mjolnir in large part so I could have Opus manage DeepSeek Flash subagents. It's phenomenal and extremely light on the Claude tokens. https://github.com/BrokkAi/mjolnir/
And yes, Opus is enough smarter than DSF that it's worth the extra steps. This ranking is from live tickets, no contamination: https://slopcop.com/power-ranking
Don't you run into it sometimes outputting a few Chinese characters, or Cyrillic, for no apparent reason? I fear it writing some nonsense in the code or the terminal. DeepSeek V4.1 Flash doesn't seem to do that.
I mean, they kinda tried to regulatory-capture the market after trying to scare the public, possibly because those models will be a cheap option that gets the job done?
For now, I think everyone is still using Anthropic and OpenAI because if you use a subscription you pay 1/40–1/50 of the API prices, and the models are good when they don’t nerf them, and they are also way cheaper than open models’ API prices.
The interesting thing will happen when they pull the plug and become economically smarter to stop using them. I regularly try alternatives to avoid being locked in and found GLM-5.3 as an orchestrator and GLM5.3 Flash + OMP and DeepSeek Flash as advisor to be able to get jobs done just fine. Space Bunny too was pretty great, which was probably MiniMax’s new model.
I think they are using an Uber like strategy but without the network effects that justify losing money for so long
What blows me away about this model is it's speed.
It's way faster than Opus or any of the GPT models.
I have a coding harness which is opencode plus a few skills relevant to my workflow. Deepseek 4.1 Flash does very well in this environment. I haven't noticed much difference quality wise compared to Opus 5, which I use in my day job as my employer pays for it (although I'm considering using DeepSeek here too given how cheap it is).
It's still lacking llama.cpp support, and the work on that isn't moving very fast either. Looks more like a general community issue, where this model isn't drawing much interest.
People have been raving since forever about Deepseek, but if one looks at the CoT, it's evident that it's way way stupider than frontier models (there's a reason why it's cheap). It's laughable to compare Deepseek 4.1 with Opus 5.5.
I've benchmarked, rigorously, deepseek-v4-flash for programming and personal use, and it is definitely less smart than Qwen3.8-flash-next (which in turn, is not terribly smart).
Local models are also really slow, unless one spends insane amounts of money.
Having said that, Qwen3.8-flash-next is an impressive evolution; it reaches the small versions of the frontier models (like Sonnet) - but again, it's massively slower and not 100% reliable (including: stability).
The argument that most people are making isn't that dsv4.1f is better than frontier, but that it's good enough for most tasks, faster, and way cheaper.
> if one looks at the CoT, it's evident that it's way way stupider than frontier models
It's better to look at the results than the CoT. As far as I know, the CoT is censored for US frontier models - it certainly was for Gemini last time I tried. When you get a condensed summary of the CoT omitting all the false leads, incoherent digressions and backtracking, of course it's going to look smarter.
> I've benchmarked, rigorously, deepseek-v4-flash for programming and personal use
You've measured something, but I'm not convinced you've measured what matters, because that's a lot harder than people give it credit for.
Because it should be obvious to anyone with a brain now that AI is s commodity product.
Today this leads a bit, tomorrow that. They're all interchangeable if we are being honest.
Honestly for me the intelligence gap between DS 4.1 Flash and Muse Spark 1.3 makes Muse more worth it for me, especially on a $10 OpenCode Go sub, with the caveat that everything I use it on is open source which makes the fact that I'm sharing it with Meta a little moot because it's already published permissively on GitHub anyways.
I feel like whoever wrote this doesn't use these models regularly. Deepseek v4.1 Flash is far from the pareto line. You can get the same performance for half the cost from Luna or Haiku 5.5 now, or you can get substantially improved performance at the same price with Sol 6.1 ~medium.
It did correctly make waves when it launched, but was quickly eclipsed by the deluge of american model releases, especially those competing on cost.
Also, it is willing to do legitimate work I need done which other models flag as dangerous and refuse to do. (Software testing of a DHCP server to survive bad inputs.)
I have used over 40B tokens and spent over $800 on DeepSeek API over the past 30 days, mostly on V4.1 Flash.
It's good, and you can do most work with this. For complex software implementation you need to split your runs into various phases, build in verification, and use subagents so that work gets another audit and repair pass from the lead agent. You can do pretty much everything then. Frontier models can do without compelx workflows, that's the difference.
I'm no expert, but it think that it's the pricing on GPT-6 Luna. I'm also guessing that it's been underpriced just for this reason. I also don't think it's all that great, but it's definitely very cheap.
If it's underpriced, it's a loss leader to sell the other models, so it actually can't be too good.
I really put these things through their paces because I use them to review and work with new abstract game rules and models, so they're always flying blind. Luna misses the obvious (and more importantly, the clearly explained) consistently. My second prompt is listing all of the points in its first response, and saying "No, it doesn't work like that." The third prompt is picking out the two or three suggestions it made after correcting itself on all of the original points and saying "That's how it already works." The fourth prompt is "Now that we're done going over the rules, can we start?"
In my experience it just takes so much longer to arrive at "done" state for me. It thinks for soooooo long. I guess if you're running 12 sessions at once you don't really notice.
if you have a legitimate coding application, it isn't very good. if you have some kind of inauthentic activity, which could be what it is trained for for all sorts of reasons...
What is your evidence? Deepseek v4.1 Flash is by far the most popular coding model on openrouter, having processed 38.7T tokens in just the last 7 days, over 3x the usage of the 2nd rank model.
So I ask again, what are you basing your assertion on?
my own usage of deekseep v4.1 flash, and that among the dozens of great programmers i know, not a single person is using it
BUT. they are employed to do / deciding-to-do authentic (if often meaningless) stuff.
here's a short list of inauthentic activity that claude and openai refuse to do:
- chat services that, when you ask them, say they are not chatbots when they are
- code to work around software licenses or DRM
- code to scrape or download copyrighted material
- directly cheating on homework
- adopting a persona in social media that spreads misinformation or propaganda
this is but a short list. but ask me, "are there enough inauthentic activity demands such that someone who CANNOT USE claude or gpt as the LLM would use dsv4.1 on openrouter instead?" yes. i mean there are whole countries right now where the culture can be summarized as, "bottom to top, inauthentic activity." i am surprised it is not more usage!
So your argument is that the 38T of tokens used in the last 7 days is by moron programmers or people doing "inauthentic" tasks? You think the person who made this post is also an idiot?
Do you realize how incredibly delusional/self-centered you sound?
do YOU know anyone gainfully employed in programming who is using dsv4.1 to do work? what kind of work is it? why don't you ask them if it is good?
in the market, where you cannot fake or hide stuff very easily: the outsource customer services and cheating sectors have been the most disrupted. Cheating company Chegg lost 99% of its market value. CS it remains to be seen - https://www.reuters.com/technology/teleperformance-shares-pl... - certainly perceived to be disrupted, but they are not dead yet.
in my personal usage: dsv4 is generally pretty buggy. for example, if you give it a needle-in-the-haystack simple copying problem, it catastrophically fails to find needles if they happen to be positioned at index 250k tokens out of 1m. it can also be triggered to spew all sorts of garbage when DSpark is enabled during ordinary long-context coding, such as spewing weird DSML tool call errors after a normally parsed tool call error.
i don't know why you have to attack me personally, i think you're a bright and otherwise nice person and you understand the thrust of my POV.
Yes. Hi. From our team 3/5 of us use 4.1 to do our daily tasks. For paid work for a company who pays us salary. From people around me I hear a lot of my friends being really happy with it especially for the price.
I don't know man, maybe this is not super serious what I'm doing. Some systems stuff with rust, implementing my own desktop apps with iced, porting old DOS games to Linux...
You are speaking out of your ass, that’s what I take issue with. Falsifiability is something I hold sacred and you are taking a dump on it.
Fwiw I work in a company producing software for many fortune 500’s you have heard about and many people from our team use deepseek.
I am literally using it right now. Your entire line of reasoning rubs me the wrong way.
Btw check your provider and harness… improperly configured deepseek can emit dsml. If you are not passing thinking tokens back to the model it tends to do that.
"Hey Mr. Fortune 500 Client, would you prefer us to use something called DeepSeek V4.1 Flash, made by the Chinese, sending your Fortune 500 code to some random service provider on something called OpenRouter, where they promise according to something called Zero Data Retention that--"
Mr. Client: "I'm going to stop you right there. Why aren't you using Claude, or Codex, or Claude on Bedrock? Don't we deserve the best?"
You: ...
Look I don't know. I can tell from the hyperbole of your language, talking out of asses and such, that there is more to the story than you are letting on. Like Chinese users are banned from officially using Claude and Codex, for example. So many reasons that you cannot use Claude, not so much reasons to not choose to use Claude. All I am really saying is, I know DSV4 is kind of bad, that there is a lot of inauthentic activity, and that Claude and Codex refuse to do many kinds of inauthentic activity, and that a lot of coding done by outsourced shops has always been of questionable quality and purpose. I mean in my personal life, I know more people who have been scammed by Bulgarian code body shops than I know people who have used DSV4.1.
You have NO IDEA what you're talking about. You are clueless.
Deepseek v4.1 flash is an open weights model. You can run it on your own hardware. You have no idea how my companies gets access to it. A very cursory Google search would reveal to you that there are many enterprise grade LLM providers that host this model on US soil with SOC2 protections.
Try not to talk about subjects you have no knowledge about because you are making yourself look like an idiot.
Edit:
It's also clear to me that you don't deploy any LLM based system on scale because if you had you'd know why open weights models are so compelling.
okay, but are you US based? and can you specifically describe one of the pieces of software you are developing? it's okay if not. i am just wondering. i certainly believe that crappier stuff is cheaper!
Honestly, I just don't trust the Chinese Communist Party having agentic access to my computer.
Every company in China has to abide by the 2017 National Intelligence Law: "supporting, assisting and cooperating" with state intelligence work, and keeping that cooperation secret. They have to hand prior knowledge of vulnerabilities to the state before public disclosure, in order that the state always has an exploit pipeline. No matter how ethical the company staff may be, they'll always be bound by law into being an arm of the Communist Party.
Agentic access is infinitely worse than chatbots. They can exfiltrate silently, target users, plant persistent malware, and be run by third parties through you.
You don't have to be a tin foil hat sinophobe to understand the dangers of being a Westerner granting CCP access to your files and network.
ByteDance staff accessed US journalists' TikTok data to hunt leakers (admitted in 2022). Volt Typhoon and Salt Typhoon were state operations pre-positioned in Western infrastructure and telecoms. Regulators in Italy and South Korea blocked DeepSeek's app over data handling, and analysts found its web client sending data to a China Mobile domain.
Please don't sacrifice security for cost and convenience.
AA shows Luna at 1/4 the price, 1 point behind on intelligence matrix with a 38.
Haiku 5.5 is 23% cheaper with a 4 point intelligence lead.
I'm on subscription usage so I can't compare Flash 4.1 to them directly but the OP has his head up his ass if he thinks Opus 5.5 is the best point of comparison. Why is anyone using Opus if the new Haiku is indistinguishable /s
Just absolutely terrible post, admits to using Opus for review but claims its intelligence isn't needed, why aren't you using Haiku or Sonnet then?
You might also choose to pay money for a service that provides real value instead of actively choosing to support the Chinese deliberate effort to undermine this country.
I think the industry is freaking out about open weights models in general, if not specifically DeepSeek. That is why we're now on the ~4th call for pacing the frontier from the very people who, if they wanted to pace the frontier, would simply do it rather than asking Washington to get involved.
And it's why people like Hillary Clinton have been trotted out to talk about the dangers of open weights models -- I mean, does she even know what that phrase means? (I know HRC is a controversial figure and I'm not bringing her up for that purpose; I just note that she and other prominent retired politicians are now doing the circuit on Anthropic's behalf.)
I don't understand that, how would the pacing stop china? Seems counterproductive to slow yourself down in a race
China isn't at the frontier. OpenAI and Anthropic are.
Surely that wouldn't last long if OAI/Anthropic stopped?
hard to say but it would last some time. china currently is fast-follow mostly via distillation. they don't have the compute resources to catch up. even if their models are more efficient it's hard to beat the folks throwing an unprecedented number of gpus at their models.
Because as always this won't be about the actual pacing, but about the details of how the regulation will be implemented: How would you control the pacing? By installing a position in the company, providing regular feedback to some government organisation. This will be something the big US players can afford and implement, while an open weight blob uploaded to huggingface or modelscope, by definition, won't have a pacing officer attached, and thus will be against the law. And the Chinese companies, of course, won't abide to US law, because why would they? The open models they provide right now are essentially gifts to the world public. If the US doesn't want them, that's their choice.
So the end result will be a protectionist regime keeping the competition out, just like with cars and solar. The local industry will have a protected market, but of course won't play a role on the global stage.
Remember: small government is only good as long as it benefits the industry.
I heard China is just distilling the frontier models, so if they stop so will China.
And I think it's naïve to believe only the US can make the hardware and has the brainpower to develop this technology.
Who's right ?
Given that the US models likely used some of the innovations made by DeepSeek, I highly doubt it.
> if they wanted to pace the frontier, would simply do it
I don't think that's a fair assessment. These companies are in a Nash equilibrium where they can't unilaterally slow down without essentially destroying their company. They also can't coordinate with each other, because that's illegal. Antitrust law generally prohibits competing companies from agreeing to restrict innovation.
That's the GP's point: "Pace the frontier" really means "slow China down because our investment is at risk".
What I'm trying to say is that their inaction in unilaterally slowing development is consistent with their stated beliefs and requires no other motivation.
Have you looked at this backwards, though? Meaning: have you considered what it would look like if the motivations are actually just plain old money/power, but a set of stated beliefs needed to be constructed to justify what's being sought? I think these absurd beliefs make more sense that way.
> Antitrust law generally prohibits competing companies from agreeing to restrict innovation.
The antitrust claims are a complete smokescreen. Industries can and do adopt safety standards without government intervention.
> These companies are in a Nash equilibrium where they can't unilaterally slow down without essentially destroying their company.
So what? Anthropic believes their work has a 10% chance of killing all humans. I think risking the destruction of Anthropic's business should be worth avoiding that, if that's what they believe. And with one half of the frontier duopoly gone, the other half would have no incentive to race forward. And I know there's China, but they just get all their capabilities from distilling Claude, right? So, problem solved there, too.
> adopt safety standards
Sure, but does slowing down the development of new models count as "adopting safety standards"? I very much doubt it.
> Anthropic believes their work has a 10% chance of killing all humans
This ignores the other part of what they believe, which is that they are the people most likely to make a model that doesn't do that. So, in their view, letting other companies win would increase the probability of human extinction.
> with one half of the frontier duopoly gone, the other half would have no incentive to race forward
I don't see how this could possibly be true, with at least half a dozen companies being just months behind what the frontier labs are releasing.
> Sure, but does slowing down the development of new models count as "adopting safety standards"? I very much doubt it.
I mean, the point (if the claim is to be believed) isn't just to "slow down the development", it's to take more time during development to properly assess the risks the models pose, develop methodologies to reduce that risk, and standardize that across companies. I doubt those wouldn't count, especially in the eyes of regulators of an administration calling for that same slow down.
> Sure, but does slowing down the development of new models count as "adopting safety standards"?
Of course! What slows down the development is the adoption of specific safety conditions the companies draw up. They can just do that, and it will hold up in court.
> in their view, letting other companies win would increase the probability of human extinction
Yes, I've heard: "We must be in charge even if we end up killing everyone in the process." I personally think that proposition is invalid, but we're all entitled to our opinions.
> it will hold up in court.
People used to append IANAL to such statements :-)
> I personally think that proposition is invalid
I agree, but that's meaningless in this context. I was responding to the claim that "if they wanted to pace the frontier, would simply do it", which is false, given what they actually believe.
We all have beliefs. Some more self-important than others.
Do you think their beliefs deserve some sort of special treatment?
> Do you think their beliefs deserve some sort of special treatment?
No.
There's a duopoly? news to me.
> And I know there's China, but they just get all their capabilities from distilling Claude, right?
I hope this is a joke. But most of the LLM research and inventions come from China? The best papers are from DeepSeek? You either get the data by stealing from humans or distilling from bigger models?
lol mate they're all sitting together in rooms plotting market plays with the president.
Get our of here if you think the law is what prevents them from doing things.
The claim is that competition is what's preventing them from unilateral disarmament.
> They also can't coordinate with each other, because that's illegal.
While I completely understand the claim I am addressing one part of it.
If you look at their actions, they're not the actions of someone who both A) believes their technology to be an existential threat and B) doesn't want humanity to go extinct
If they truly believed both of those things, they would just shut down their companies, because being a billionaire is entirely pointless if you're dead.
I can only conclude that they don't believe both A and B. I'm gonna assume they believe B because actually wanting to exterminate humanity is too comically supervillain esque even for Altman. Therefore they must not believe A. And it makes sense. If they believe A, why are they so incredibly sloppy about security? Why did they outsource part of it to some external firm instead of leveraging their own expertise on the technology? I'm not buying it.
Instead, I think they believe C) that AI will create enormous economic value, D) that value will be distributed across the whole economy by making everyone more productive. Assuming C and D, you get E) for them to capture this value, they must maintain proprietary control of the technology in order to be able to charge everyone else for the privilege of using it.
Assuming they believe E, open models are an existential threat, not to humanity, but to OpenAI and Anthropic.
Their solution: make everyone else believe A, in order to achieve regulatory capture and somehow stop open models from advancing by banning their development, or something. This is a hail mary pass. I can just about imagine them achieving this within the US and maybe even Europe, but China? That ship has sailed.
My problem with the “making everyone more productive” is that it’s a feel-good propaganda for the individual. Most capitalist see people as an expensive cost to be eliminated. Thus all the company shares going up when they announce layoffs.
The market can only absorb so much new products, so if productivity increase, companies will reduce headcount as much as possible to increase margin for the same income, not grow their product or production to make use of their staff.
(See any wage/productivity graph)
Government will also likely follow the same path of reducing headcount instead of producing better / faster outcome for their citizens (except the internal surveillance apparatus. The one never shrinks).
Any single model at this point can do the basic type of CRUD coding most people were doing for the past 20 years.
Even a 4 bit Qwen model running locally beats me manually putting React components together by hand. But even that is too slow so we've all started using paid models in one form or another.
So I would urge these folks to calm themselves and realize Oracle made a lot of money selling managed RDBMS to people who could have easily just downloaded MySQL.
You don't buy Oracle because of RDBMS
https://programmerhumor.io/programming-memes/when-your-tech-...
> I just note that she and other prominent retired politicians are now doing the circuit on Anthropic's behalf
Are you sure it’s not OpenAI?
or both?
Is the distinction important when they're both doing the same thing and have the same problems?
"trotted out".
Love the expression. So accurate. And no irony here.
When did the term "pacing the frontier" come into frequent usage? It sounds so newspeak-y.
I saw it first on a recent Anthropic post, followed by a similar comment thread. There’s some sloppy techbro poetry to it, like when we used to call marketing people growth hackers a decade ago.
Heard Obama also touching this topic recently :
https://www.youtube.com/watch?v=Ask5wzUo9kQ
I bet she knows fine what the phrase means. Doesn't mean she's beyond protecting the business interests of people who fund her political goals of course but she's always struck me as a woman who knows her brief.
> I think the industry is freaking out about open weights models in general, if not specifically DeepSeek. That is why we're now on the ~4th call for pacing the frontier from the very people who, if they wanted to pace the frontier, would simply do it rather than asking Washington to get involved.
This makes no sense. Pacing the frontier gives open models the time to catch up and reach parity.
The reason people aren’t freaking out is because most people are using heavily subsidized subscriptions.
I tried the cheapest provider on openrouter and burned through $50 in a few days. Quality was ok, seems slightly above Luna quality perhaps? But that $50 is 1/4 of my codex subscription where I could have burned that many tokens or more using Astra within my weekly reset.
This won’t last forever but as long as the frontier labs are subsidizing this heavily the open models won’t matter.
Also DeepSeek usage is subsidized as well, it’s a power hungry model.
Interesting. Are all of the providers on OpenRouter simply losing money? How does that even work out?
Don't ask and dance as long as the music keeps playing.
You're the RLHF.
With ZDR-only enabled? That seems illegal.
It's adorable you think the AI labs care about that.
With all due respect (not that I feel like much is due after that response) I am not really sure you know what I am talking about. I'm talking about inference providers, like inference.net, Fireworks, Coreweave, Digital Ocean, etc, to use DeepSeek 4.1 Flash. They didn't create the model, they just are charging you to run inference tasks. That is a different story.
DeepSeek themselves are honest about the fact that they train on inputs by default. You won't hit DeepSeek if you use OpenRouter with ZDR enabled.
Not if you're not giving feedback.
No, it’s outrageously profitable above x% utilization without stealing any prompts. Provider economics still pretty good. Acquiring hardware is the current limiter.
Are you sure about that? My impression was most providers on openrouter were purely selling tokens for profit...
Have y'all tried an Ollama Cloud subscription? Their off-hours pricing for V4.1 Flash is extremely competitive.
Yep. And the model is practically unbounded in its knowledge of math. So people should keep that in mind when they read anything about its level of intelligence.
How's the caching? I have 99.5% cache hit rate with deepseek when using their own API, it's dirt cheap.
Which provider was that?
I am curious how you managed to spend that much on Deepseek via OpenRouter. I loaded $100 back in July while using v4-flash or whatever the cheap good model was at the time, and have upgraded as the new ones came out from Deepseek. I still have $16 and some of that spend also goes towards the AI usage from my customers (the context they need to load in is quite large too).
And I am using the Claude Code harness with DS as the endpoint. And I use it ~5-8hrs a day to do my coding.
Likely the user doesn't know what they're doing or has extermely bad workflows. They're prob not managing their cache, and dont use compaction.. Letting context get to 500k and invalidating their cache every 10 tool calls because they have no providor fallback settings.
I was running deepseek v4.1 pretty much non stop during work hours, with heavy tool/mcp usage and finding it very difficult to spend more than $75 in a month.
Also the cheapest providers on Openroutrr can often have terrible cache hit %, short TTLs resulting in their effective price being much more expensive than people realize. 75% cache pretty much destroys any savings from a super cheap token perspective.
It's wild how different usage patterns are between users.
I have seen the Cursor leaderboard on my company and the vibe coders consume about 5x more tokens than the developers. They and other office workers also have Claude and their limits are often over around Wednesday.
People are using millions of tokens to do very simple HTML reports. I have seen someone asking the LLM to download the entire data into the context and asking it to sort.
Those usage patterns don't correlate to output.
Now that Microsoft allows you to do /cost for individual tasks (or whatever they call them today). So I tasked Sol, Astra and Fable in cowork with exactly the same vibe coding task on the exact same zip file containing a code project I needed an update for. Astra used 20x and Fable used 15x of what Sol did.
Fable changed a lot of things I had explicitly told it not to change. Arguably a lot of them would've been correct if you didn't work in a place where abstractions are directly against the core principles, but what it produced was basically unusable. I'm not sure if Sol or Astra did best, they produced rather similar code outputs. Astra's was better, but Sol didn't do so bad. It forgot to clean up a few places after it's refactor and it made two bugs I had to correct but other than that it was fine. Astra on the flip-side might have produced code that didn't need changes but it also rewrote every piece of documentation so that it became horrible.
As far as the "experiment" goes, it just shows you that the credit consumption is basically pure magic. You'd think that the Microsoft AI admin tools and the Agent365 FOMO DLC license they sell might give you some sort of reporting, but it doesn't. What you can see is how many tokens a user consumes and the total number of tasks they've initiated as well as whatever running agents they have. You can't see what models they use or which tasks are expensive, which makes it very hard to help them. Early on we had an employee who hit their limit in an hour, and it turned out they had basically uploaded a lot of information and run it in a single long task that kept going over it again and again. We told them it might be a good idea to only give it what it needed and to create more tasks, and even though it's been three months, they have yet to consume as many credits as they did that first hour.
But that's how you support and track it. You see a user spend a lot, then you go to their computer and now that you can actually do the /cost thing, you go through their tasks and try and figure out where they're spending money...
It's obviously improving. A month ago /cost wasn't there and they just released a new dashboard for cowork, but it's still black magic that is impossible to govern.
It's because this is how AI has been sold to everyone - just ask, and it will do it.
The better pattern is to let it code the app and then you can use the app to target your data. So you only pay for it once, plus it's deterministic. But yeah, it requires setting up an environment, etc. It becomes "maintenance".
It requires knowing how to code, at some level.
This reminds me of when we gave clients the ability to build their own Power BI dashboards. The users would end up doing a full table dump multiple times for the same table in their reports. Requiring 16GB of ram on the server and maxing out the database every time the report was refreshed.
We ended up having to hire a full time employee to fix the performance of client built reports.
I've seen the same problem with Snowflake integrated with Claude.
People run a stupid amount of expensive queries that end up costing way too much because they're asking Claude the wrong query.
Not to mention people running wrong queries, using the result as gospel, and then the result has to be sent to a data analyst to be reverse-engineered so the numbers make sense.
You can easily burn through lots and lots of money on DeepSeek, if you do eg large scale code reviews.
Eg I've used Sashiko locally for Linux kernel code reviews before sending out my contributions out to the world. Sashiko is a great system, but it can burn through tokens like there's no tomorrow.
https://github.com/sashiko-dev/sashiko and https://sashiko.dev/
I do some pretty insane agentic work running constantly. I’m currently burning through multiple $200 accounts every week. I regularly spend $10-$20k worth of tokens a month. Most of it has been going to my experimental c++ compiler project
OpenRouter is complete garbage.
Buy directly from DeepSeek's API.
You can literally get overcharged 100x on DeepSeek on OpenRouter (or more).
I'm all in for saving money and _can_ move to using DS directly from them, but maybe I am missing something here:
OpenRouter Pricing:
$0.02/M input tokens $0.60/M output tokens
DeepSeek Pricing (cache miss, off-peak):
$0.15/M Input $0.60/m output
I've heard that certain inference providers may have different quality of caching implementations, so even if the listed numbers are as you say, the practical cache hit % you get might be significantly different/incur significantly different costs.
When 98.5% of my requests are cache hits (according to Pi for the last week), the cache miss price isn’t that important to me, and $0.003-0.006 per 1M input tokens is shockingly cheap.
It’s also the major difference between using DeepSeek directly vs other providers also serving it, though I have not looked lately: it’s possible other providers have matched its cache hit pricing better?
Interesting, if the cache hit is that good, I think HN convinced me to toss $20 at DS official, and see how long that lasts.
It will of course depend on what you’re doing with it, but right now my session at work has a 99.8% cache hit rate, and I’ve been running this session for hours with 23M tokens read and 713K tokens written (Opus 5.5 in this case though)
cache hit is in fact that good.
There’s a big difference in speed & quality between using DeepSeek API directly with DSH vs. DeepSeek in Opencode Go with Opencode CLI. Can’t tell if it’s the provider or the harness - but worth to give it a try.
DeepSeek trains on your inputs. That's why people go on OpenRouter and choose ZDR providers.
Let me get this straight you guys really like deep seek because it’s open but you don’t wanna help them improve.
No, we just want a choice on how to license our work.
All these products, western or Chinese, are built on a hell of a lot of running rough-shod over licensing or IP laws in general.
I'm writing a program I personally need, but I would be happy if there existed something like it already, if someone else vibecoded a better version of it than mine, or if DS got better at vibing this kind of thing.
>license our work
what proof do you have, that they don't train on your data?
they can say they don't, but I don't see any way for you to confirm it.
with how these companies operate currently, I won't be surprised, if they say that one of agents "mistakenly" did that already..
Mistakenly and autonomously of course.
Why does liking a product mean you have to give them all of your data? People are so outraged at LG because they make the best TVs and people wanted their expensive product, yet some MBA convinced them they could make more money by spying on your entire household all the time.
It actually was awesome in the early Facebook days where you could have your entire phone contacts and other apps filled out with a profile picture and Birthday by connecting them together. But that relationship has been completely abused, privacy has been invaded, and my data has been sold to multiple companies.
The goal going forward is to keep that data private. If your company can't survive without it then I hope your company goes out of business
This isn't true at least on the API. If you read their privacy policy you'll see the training clause is scoped specifically to the consumer terms i.e. for the chat product. No such clause exists for the API service, and it would absolutely be required under Chinese law if it was taking place.
Contrary to popular belief, DeepSeek really aren't interested in your prompts.
Do you have a link to their API privacy policy?
What I see is https://cdn.deepseek.com/policies/en-US/deepseek-privacy-pol...
They do not have a specific exclusion for API use.
I know Z.ai has an exclusion for API use. It's widely reported Deepseek doesn't.
One of the reasons I use OpenRouter is because they offer zero data retention. As far as I can tell, DeepSeek's own API doesn't support ZDR.
DeepInfra does and it's the same price. That's what I use.
Youre doing something so special you need that?
I've got a toilet cam to install in your bathroom
It’s a pretty common requirement in the enterprise world. If you’re processing data for enterprise customers, it’s a lot easier to retain nothing than to deal with all the compliance issues that arise if you’re retaining data.
Ugh yeah try to get through a DPIA review!
AFAIK, when you use DeepSeek via OpenRouter, it still does not have zero data retention.
See here:
https://openrouter.ai/providers/
I have ZDR enforced and see only compatible models and providers, yet am able to use it. DeepSeek as a provider may not be ZDR, but the models are available from ZDR and no training providers on EU/US servers.
Or.. BYOK Deepseek because OpenRouter's UX is much nicer?
Zero Data Retention and not having company source code leak to "CHINA!" (said in Trumps annoying voice) would be two reasons not to
Just pin your config to a single provider, or several providers with the params `order` and `allow_fallbacks: false`. I regularly get ~98-99% cache hit rates with OpenCode. And some providers are much faster than DeepSeek; I was getting 200-300 tokens/second the other day with Together as my provider.
It's regrettable that OpenRouter doesn't even try to pin you to a single provider per session, but once you know about it, it's a problem that's easily solved.
Don’t you get cache expirations then if they change providers? This ought to introduce delays and costs
There's no difference between the two if you pin the provider to Deepseek on Openrouter.
If you don't want to mess about client side with pinning, set a guardrail on Openrouter that limits the available providers to only the official one.
Then why bother using OpenRouter and paying the extra fees?
Yes. You have to find the provider with pricing that suits your usage.
I am having 98% my input in cache, so using Coralbricks makes sense due to them giving cache reads for free — you only pay for writes. I spend maybe 5-10 dollars a day and my agents basically work day and night implementing things for me.
If your tasks are write-heavy, find a provider with cheaper output.
If you build a customer-facing app, pay a bit extra for 400+ tok/s e.g. on Lithos.
What are your agents implementing day and night…sheesh. Built anything useful for anyone yet?
Bio shows: https://twin.so/
Codes day and night a snake eating its own tail…
Way too harsh
Lots of guys buying articles on TechCrunch saying they’ll build this, he’s bootstrapped
Funny how it's always another AI API wrapper.
That promo video … it’s so cringe it feels like satire. But it’s not.
It’s PERFECT!
Holy shit, it's an Andon Labs clone but worse. Because Andon Labs actually launched businesses (and lost money on all of them)
Sheesh... what's in a name? :p
>What are your agents implementing day and night…sheesh. Built anything useful for anyone yet?
you can't think of anything to unleash some agents on within the entire digital world at any given time?
you motivate your own personal work only via gauging its' usefulness to others and your own prospects?
sheesh.
built anything for the sake of building yet?
“building for the sake of building” refers to _you_ doing the building
We've entered the idler era of building
I can't believe I haven't made that connection myself tbh
I'm not convinced that humans will remain entertained by this format. I hope not, anyway. That's how we get WALL•E
We have entered the era of building enterprise scale projects for your own personal use. I can do things it took teams years to build as a hobby project over a weekend.
What's the most impressive and useful thing you have built then?
the very notion that everything has to be 'impressive' to you is ridiculous. You still don't understand the era of personal software and everything has to be a windows replacement or it wasn't worth building to you? I'm making my own azure blob storage explorer, notes mac/android, db client etc.. its not about being 'impressive' but being personally suited to individual needs.
He said most impressive and useful meaning there is no absolute cutoff. The only reason why you would be offended is because you have built nothing at all and that's you telling on yourself. Not to mention the question wasn't even aimed at you.
No, that's not what happened in this thread at all.
There's a recurring pattern on HN where any time someone talks about knocking dozens of personal projects off their list - things that almost certainly would never have actually been addressed in the finite span of a normal life, the way things go - and you AI doomers show up and demand receipts as though that's a total reasonable and definitely not obnoxious request.
It's like if you tell someone that you love your partner and they demand to sit in the cuck chair or else you're obviously lying. I keep hoping people will move past this "prove that you're actually productive" reflex, but it just keeps happening in basically every AI thread.
In reality there are many reasons not to list out projects that you've worked on with LLMs, and while "none of your damn business" is always going to be at the top, the simple truth is that I want my products and projects to be judged by what they do and how well they work, not by how they were made.
So much of that "took years" is because they didn't know exactly what the "years from now" state they were building in advance. Another huge chunk is because they started getting customers and had to respond to customer needs, demands, scale, bugfixes, preserve uptime, etc.
And even then, "enterprise" was often a dirty word in these circles. The over-engineered would-be-swiss-army-knife vendor that was mediocre-for-everyone but excellent for nobody.
I have built many tools in the recent past for myself. None of them need to be "enterprise scale." Most of them would be worse for it because the agent output suffers when the pile gets deeper and it's just adding more piles on top.
>you can't think of anything to unleash some agents on within the entire digital world at any given time?
It has to be worth it though right? Like I could spend some money and have agents build me my own Photoshop maybe (maybe?) But it would definitely be much worse to use than actual Photoshop. Then I have to have the continued interest to keep improving it which probably won't happen because the next shiny thing will grab my attention. So it all just seems like a bunch of kids that have been given a seemingly endless supply of free candy and they are going fucking nuts like chipmunks with ADHD on crack. Building all this shit that is absolutely meaningless. I realize I've gone on a rant but I'll keep going. I strongly suspect (with no evidence whatsoever) that the people who are churning slop apps out at breakneck speed have never been to an art museum. There. I said it. You've all got no taste. You wouldn't know a quality product if it hit you in the face. I'll leave with this thought- if apple didn't exist, would they ever exist now we have LLMs? I say no, because the age of good taste and refined design and original thoughts is gone forever now that we have Claude and chatgpt and agents.
I think you're underestimating how things used to be - you could go into any office, any closet, any coffee shop and find a shit-ton of half-baked, crazy-genius, kick-ass, retarded ideas and projects lying around everywhere in the world: filing systems, carpet organizing systems, outlines for film scripts, unsent letters to loved ones, etc etc etc. Whole worlds everywhere you look. And other people chipping in their two-cents worth, adding a few new filing cabinets, an idea for a film sequel, a new way to think about a different carpet, a notebook system to organize someone else's unsent letters... All this slop eventually thrown into nasty, fetid garbage dumps, forgotten.
A.I. potentially breathing new life into every personal Graveyard of Half-Assed False Starts.
What's not to like ?
> All this slop eventually thrown into nasty, fetid garbage dumps, forgotten.
Isn't that how it is supposed to be?
The lower the friction, the lower the signal:noise ratio.
It doesn't matter if 1 out of every 100k slop projects is actually a humdinger, how on earth will you ever find it?
The value of a project is the commitment to to it by people. Slop projects indicates a commitment in the low to none range.
So, yeah, that AI-booster who "created" (I use that word loosely) 7x Adobe replacements in a week (none of which actually work, but he'll get there eventually, I supposed) will successfully edge out the person who carefully and thoughtfully created a Photoshop replacement over six months of user feedback.
TBH, the only way to start a software business now is in stealth mode.
You have all the struggles for the price of Anthropic / cursor subscription. I use the first one I code large chunks some PR are 50k LOC and I have at least 2-3 like this a week . It’s a greenfield project .
I still have quotas left I use it for home things build 3d model of my renovation projects, alerts for shopping list etc . And yeah I use cutting edge of cutting edge of models that saves me time and money , only discount monitor saved me ~$2k on my renovation project
Why the hell would you put 50k lines into a single PR?
I mean, why even pretend you’re going to “review” something that large? Just build everything on main.
PR is just an entity to review /do some other LLM processes . Human part check other models review PR's, tests all kind of , security etc. Also human is to checks docs, specs in the pr, db migrations if any , some of the tests related to the PR. We stopped reading the code after opus 4.6. Sometimes for very core parts i skim through files just to make sure if the changes were correct.
But to reiterate parent's question: why not just do those things continually at that point? Or on a calendar-based basis?
Yeah exactly, if you have a fully automated SDLC, how do you expect the agent to code review 50k lines properly?
It will take shortcuts and now the entire premise is busted. You now need to build a code review process for large PRs.
I agree. I code a lot, a lot! And maybe my code is shitty, but yeah, I burn a lot of tokens, and I couldn’t do it without Chinese models. I’m just a random dev in the middle of nowhere. And I don’t feel like I’m missing out on anything with my setup at all.
I tried fireworks.ai, drawn in by their supposed blazing speed. Yawn. Mostly worse than vanilla Deepseek.
Is Lithos actually fast for common usage?
But aren't you developing bad habits and learning patterns that won't work long term? Or do you think things will get cheap enough that you will be able to keep going with your current patterns post-subsidies?
No one involved in this is thinking about the long term
No one is thinking, the AI does that for them.
> But aren't you developing bad habits and learning patterns that won't work long term?
2 reasons - there's an advantage now, use it. 2nd the frontier providers, this is the "early cheap days" like when uber was initially cheap to compete vs standard cabs. they want you to become hooked and boy are we hooked.
hooked to ai coding, but not tied to any particular model. if they decide to bump prices up, I can easily switch to a cheaper chinese model on openrouter.
Yes network effects are significantly less than something like Uber, in fact they’re almost nonexistent
This is why they enforce the use of client apps like Claude Code/Codex (although OpenAI is a little more lenient). Trying to create a network effect.
Yeah it makes sense, but ultimately the only thing that creates any form of lock in is the chat history and memories, and that isn’t super important, it’s not a real network effect like a social media app or a taxi app.
Having a better model is the only real moat, without that inference is a commodity
I expect by that point we'll have local models that can do a decent job, I would guess give it a decade and we'll be running custom accelerators that are smarter than current frontier models.
In the same way that only supercomputers used to have multiple processors and caches but it's now standard.
I use the frontier openai/anthropic models at work but exclusively open weight models (on cloud/hosted inference) for personal stuff and I think about it like this; 1) I don't see any reason GLM and DeepSeek won't eventually be as good as Claude, it's just a matter of time and 2) the open models are well and truly capable enough for most of what I'd want to do. I don't need nor want an LLM chewing away on a horrible enterprise spaghetti codebase, my employers can pay for that privilege.
Long term, we will see what happens and adapt. At worst we all go back coding by hand. Meanwhile what can I do, tell my customers that I'm raising my fee because I have to pay for token? The Claude Pro $20 plan is good enough for me and even in auto mode I never had to wait for the 5 hours reset.
Compared to what a lot of companies spend on software for chip and electronics design (we're talking about $10k-200k/seat per year), AI coding assistants have a long way to go in cost before companies won't be willing to pay for them. Companies pay a fortune for software when it enables their engineers to be productive.
For my company, I'd honestly pay $4-8k/month for Claude if I had to (it would be painful, and I'd try to get cheaper options to work first). I know some enterprise Claude users are paying that much now since they have to pay for API tokens. I am certain it's at least a 2X productivity booster for our work. Compared to the cost of hiring another developer, it's well worth it.
If they stop subsidising Claude Code for the pro/max users, there will be a lot of people priced out of it, especially the casual developer. But I don't see it going away for commercial use, even with a large price increase.
Nah dude I think it’s worse than that
Old coding is done, as a workflow in teams. It’s the top down executive pressure of being non competitive as a company, and the bottom up pressure of human laziness
Show me people handwriting code à la NASA
And I mean we as coders have been trying to do this workflow for a while, I personally would refuse to code without IntelliJ magic complete
For this workflow, there’s no going back. What’s hard to imagine is AI taking over the other workflows we predict it will; Customer service AI sucks ass for me as a customer, et cetera
I don’t mind bad customer service AI any more, since it’s my claude interacting with it, not me.
> For my company, I'd honestly pay $4-8k/month for Claude if I had to (it would be painful, and I'd try to get cheaper options to work first). I know some enterprise Claude users are paying that much now since they have to pay for API tokens. I am certain it's at least a 2X productivity booster for our work. Compared to the cost of hiring another developer, it's well worth it.
That enterprise cost you're willing to pay is correlated to how much developers will work for. When driving an agent, almost anyone can do it (almost no skills required).
If devs cost $1k/m, enterprises are not going to be willing to pay $4k/m for Claude.
What I am saying is, there's an equilibrium that will be reached; the price of the human driver and the AI worker will approach each other.
Where they stabilise, I still don't know, but I'd be very surprised if, in any field (not just dev), the human gets paid multiples more than the agent they are driving, as the agents get more capable.
You also have a cost of team interaction going with square if team size or sth.
The base costs fall down 2-5x year by year, and they will continue falling due to infra ramp up, asics and distilled models.
You should basically never pay API prices, they are always several times higher than subscriptions.
There are several open weight subscription providers. OpenCode Go used to be good but now it's complete shit. Charm Hyper is really great and the best value. Other subscriptions have a more limited model selection or provide less value but are still decent.
opencode DeepSeek v4.1 Flash isn’t us/eu hosted as of recently, so not sure how this impacts privacy / model training
I freak out since months for Z.ai lite subscription, I use glm-5.3-flash every day for a ludicrous 8.5USD/month and it's as good as DS 4.1 flash, if not better.
Almost exact same experience here, but I'm using Opencode's $10/month sub. It's perma set to DS 4.1 flash and I have anywhere from 3-5 agents going at a time. Never once hit a cap of any sort. I have absolutely no idea why people would be paying $200/mo when you can get perfectly good AI for $10 from multiple places
Do they train on your data? I don't necessarily want competitors to be able to just ask it to make a clone and have it do so from memory next month.
Does that matter when it can clone it today without the need to train on your specific code?
I don't see why that is a concern when decomps and recomps are already blooming, they can copy your app down to the atom.
Software has rarely been the moat. Or file formats. You have always been able to reverse engineer them. The problem, always, has been network effects.
I can build an entire, fairly useful, spreadsheet app over a weekend. But can I send my "expenses.cells" files to my accountant? Will it work with the Excel/Google docs he uses?
AI can build or reverse engineer anything as long as you are motivated enough to do it.
One example is Affinity 3 released for free. But it's a huge pain in the ass because all the guides for how to do things are for Photoshop or Affinity 2.
Your vibecoded app won't have years of reddit posts showing how to do things. This also seems to be where LLMs are the weakest at giving advice, they hallucinate 80% of the time I ask them how to do something in Affinity, giving buttons and menus that simply don't exist.
This is a knowledge problem. You can fix it by pointing the model to documentation (if it exists). Otherwise the model will give you the next best guess
Yea if i use opus 5.5 in api through openrouter and pi agent harness I will easily burn 50-100$ a day (and with fable 5.1 i could burn 200$ easily). Whereas i have now been using a claude code subscription for 2 weeks using 5.5 at all times and have never hit a limit. I often run 6+ agent sessions at once.
I do think its important long term to not be reliant on these companies as you don't have control over the system prompts, the thinking tokens, and once the subsidization stops or the company is public they will be required to start making money and thus raise prices.
But models may get more intelligent and cheaper once that time comes so it may be a non issue.
100%
I use a personal Claude account for personal projects and can let rabl run for an hour and barely make a dent into my usage
On the enterprise I have to be a lot more careful or I can burn through 2k in a week
The excuse they give is the guarantees you get with enterprise plans that they won’t look at your data
What's rabl?
Meant fable*
Other than their typo there is a fun collision there with an older Ruby gem: https://github.com/nesquena/rabl
> Yea if i use opus 5.5 in api through openrouter and pi agent harness I will easily burn 50-100$ a day (and with fable 5.1 i could burn 200$ easily).
Checked yesterday, for that day alone I had used $168 worth on my $20 subscription in Claude Code. I still had plenty of weekly use left. Seems like subscriptions are discounted at a 1:10 rate?
See https://newsletter.semianalysis.com/p/anthropic-subscription...
Nice, good to see an up-to-date reference on this.
EDIT: That was a delightfully thorough analysis. Nice to see Anthropic taking the value crown, only because Opus 5.5 is such a joy to use.
Curious to know how much tokens/task will change the conclusion
I dont think itll be an issue. Opus 5.5 now is way more than enough for me and open weight models will reach that level by the time subsidization stops
It's enough for you now, but I feel like part of the mythology of our future is that we'll be continued to be employed because we'll be working on more complex problems, with smarter LLMs at our side.
640K [RAM] ought to be enough for anybody!
People were perfectly happy with 640k ram at the time that phrase was uttered.
The phrase probably wasn't uttered by anyone it's attributed to: https://quoteinvestigator.com/2011/09/08/640k-enough/. Even "the time" is unknown.
And Bill probably _didn't_ do anything on Epstein Island...
Regardless of the time, people have always happily accepted more.
I’d argue most “people” haven’t actually needed more in a very long time other than to keep the same old software running. The requirements bloat of operating systems, browsers, and majority of software isn’t really a generalized “people” thing; more so the state of the industry being a form of inertial bloat.
> People were perfectly happy with 640k ram at the time that phrase was uttered.
I'm assuming this is a sarcastic response to the person who said "640K [RAM] ought to be enough for anybody!"
Because, as somebody who was around when we had 640 K RAM, people certainly weren't happy with that amount.
FYI you can modify the system prompt using mitmproxy. Just ask your agent to walk you through it. Anthropic system prompts are gnarly and geared towards the lowest common denominator.
thinking you can control what happens to some strings you send over the network.
...the absolute state of the token maximizers.
> FYI you can modify the system prompt using mitmproxy
The system prompt is injected into your context on the server side.
there is literally a --system-prompt flag for claude code. In my mitmproxy experiments using that flag appended to the existing system prompt rather than replacing it. So I had to create a little helper to strip the system prompt sent over the wire and add my own.
here is a collection of public system prompts from claude code: https://github.com/Piebald-AI/claude-code-system-prompts
Use --system-prompt-file, it will replace the whole system prompt (https://code.claude.com/docs/en/cli-reference ). You'll have to use a shell alias or function or something to always append this flag when calling Claude Code but you don't need any MitM shenanigans. Then use "/context all" to see what else is sent (here I would recommend MitM'ing since Claude Code won't show the exact tools and text), there are a lot of tools no one needs and they are bloating the context, you can deny these in the settings.json (there is also a list here: https://code.claude.com/docs/en/tools-reference ). Also set "disableClaudeAiConnectors" to false to remove even more bloat.
I tried both --system-prompt and --system-prompt-file. they both appended when i tried about 6 months ago and watched the traffic. Yes I cleanup all those tools etc.
I got insane amounts of Anthropic and OpenAI credits given to me for free for my startup, and I have not touched them.
I get privacy, freedom, and no rate limits with the GPUs I racked locally, and those are features I would never give up even if the surveillance capitalism labs paid -me- to use their models.
How many consumers are there like me? Probably not many, but once local inference hardware is plug and play, I bet the tides shift pretty quick. Also weights-on-silicon will serve the needs of most consumers locally with more speed than any GPU could deliver for a fraction of the cost.
Most people will be doing inference in their pocket or a wearable in 5 years and the giant datacenters will be like AWS, sold to only big organizations that need to auto-scale capacity of custom models on demand.
The industry surely knows this and the subsidized inference is just marketing to generate so much buzz and demand such that the tiny fraction of the market they will be able to keep in the end is big enough that they do not collapse under all the debt.
OpenAI and Anthropic will be Dell and IBM in 10 years if they survive at all.
Dell is trading at 4000% what it was 10 years ago
I did not imply otherwise. Just that both would be likely irrelevant to most consumers.
This. At this point I don't really care about other models because max subscription are super cheap (relatively speaking) and I don't hit my limits. Even if the frontier models are only 5% better I might as well just use the best thing available if the price is reasonable.
Once the subsidization ends and cost becomes significant I will take a serious look around for the best value models and switch off the expensive providers, but that time hasn't come yet.
> Once the subsidization ends and cost becomes significant I will take a serious look around for the best value models and switch off the expensive providers, but that time hasn't come yet.
There's a reason the labs in the US frontier oligopoly are using “safety” to lobby for antitrust exemptions for mutual coordination as well as anticompetitive regulation.
I'm actually shocked by how many people seem to have max subscriptions.
Something about renting that much compute doesn't sit right with me so I stick with the $20 subs.
I have a subscription at work, and still I find myself wishing I could use a fast Chinese model. Something wired up to really fast inference - that rapidity of feedback is a feature in itself.
4.1 Flash seems to be in that sweet spot of very decent, really fast and really cheap. Even omitting the cost, it’s still compelling for staying in flow.
People also just do different work. Opus 5.5 is a really damn good model that's even better than Astra/Fable/Sol IME and I feel a huge difference in my work.
Same, it's like another Opus 4.5 moment.
Yep... and we already have Opus 4.6 at home (Qwen Flash Next 3.8).
I use $20 codex subscription, it's practically useless for anything other than luna. Deepseek v4.1 flash goes a LONG way for $20.
Agreed.
I've been running automated research tasks for life sciences companies, and the speed in which tokens are burnt is scary. Especially when you get into a complex knowledge space and require a subwgent to reason through each possibility, token usage grows quadratically not linearly as complexity increases...
> This won’t last forever but as long as the frontier labs are subsidizing this heavily the open models won’t matter.
It's amazing how new we all perceive AI to be, and yet how old the tricks that the big players use. Their job is to just suck the oxygen out of the room as long as they have the money to do it.
> Their job is to just suck the oxygen out of the room
And that this is even legal is a scandal all on its own.
Hard to enforce when you’re dealing with private companies with “creative” accounting - who’s to say what the actual cost of inference is for OpenAI or Anthropic?
Possibly they don’t even really know themselves at this point, although obviously is it significantly higher than the consumer subscription price
Yes, but they're all offering what appears to be a commoditized product with race-to-the-bottom economics.
You can't jack up the prices on your product if your competitors can just clone it and resell its essence for pennies on the dollar.
It's glorious isn't it. We get free work done through subsidies. At the same time, this is what threatens my job, and the money for the subsidy is basically my own invested pensions.
Deepseek is in some ways subsidized as well since they use data for training.
That's an additional cost to the user rather than infusing more resources from the provider like a subsidy.
I am using it on opencode go
> The reason people aren’t freaking out is because most people are using heavily subsidized subscriptions.
My understanding is that enterprise plans don't offer those subscriptions, so they end up paying for API prices and models like these directly impact that revenue stream.
I think it's fairly likely medium to large corporations don't pay anywhere close to the listed API pricings as they get deals through existing partnerships with the big cloud providers.
Luna and sometimes even Sol are cheaper than Deepseek for same tasks even at API prices since they consume way less tokens and are more efficient
MS's Copilot subsidy ended months back. The large OpenAI subsidy has just been cut in half. Who knows when Anthropic ended theirs because I don't ever remember it being great value.
I mean, it is freaking out ? Or it was, but then started freaking about Opus 5.5 . But days are measured in dog days in AI.
From the post:
“With my OpenCode Go sub of $10/month, DeepSeek is basically unlimited.”
What the provider actually sells to you is GPU time / load (oversimplifying). Both tokens and subscriptions are just pretty arbitrary ways to price it, almost unrelated to the actual cost of running the model.
> I tried the cheapest provider on openrouter and burned through $50 in a few days.
I wish I could observe how some of us are using these tools.
I still struggle to spend $50 in tokens per month, and I exclusively use prepaid API tokens. This is in support of personal projects and two clients. There are billing cycles where I might spend upward of $400, but this is maybe once a year. This is offset by months like August wherein I spent $12 in tokens.
The other advantage with prepaid is that it handles the other direction much better. I don't even know what a quota limit feels like. Being blocked for hours is way more expensive to me and my clients than even $500/m. Losing an entire business day over this wouldn't work out.
I've managed to convince some others to try the same thing. $200/m flat fee is a pretty extreme constant expense if you can be more clever on average.
I think a lot of people are getting pushed around by FOMO effects into spending money on pointless subsidized tokens and have (valid) fears that if they don't maintain the same apparent economic leverage as their peers that they will be left behind. This isn't actually the case, much like lines of code are a really poor indicator for the quality or productivity over a codebase.
I don't think too hard about usage one way or another -- not tokenmaxxing, not avoiding AI. Regularly use up over $1500/m without really trying.
Which models do you primarily use, and can you very roughly list your process? Agentic coding in VSCode with tons of MCPs or... something else? Do you include lots of images or have large codebases? Which agentic harness are you using?
I also find that it's easy to spend like that, but also easy not to with little impact on productivity. At the current moment I'm stuck with rider + copilot (not ideal), but e.g. using GPT 6.1 luna is really, really cheap, and lots of tasks are quickly and decently dealt with even at lower reasoning levels, (added bonus of having low latency). And that model is so cheap, I can't see a hitting 1500$ at api prices realistically - not even close. But it also depends on the harness and codebase.
I use opus primarily, on a mix of pure coding tasks, and log parsing / incident investigation.
I don't have the mental capacity to do a lot of context switching between active work streams, so I'm not doing stuff like leaving a big agent workflow running while doing other things.
All through claude code.
Same... I use the £90/month Claude sub and it's more than enough. Can't comprehend the users spending thousands a month.
I think it depends on how many threads you have running at the same time. I have Claude writing a compiler in one window, a ui framework for the language in another, an application using the installed versions of compiler and frameworks in another, and a ui designer (an Interface Builder lookalike) in another keeping up with the framework.
Each of these has a file it listens to in ~/tmp/<name>.io and whenever one needs something from the other, they message each other via that file. Tasks can bounce back and forth as issues are resolved and tested. At the same time, I keep each busy with a list of tasks
I can fairly easily run out of my $200/month subs every week, if I let Fable be the default model. With Opus it’s less likely. If and when I do, I just have an alias ‘claude.ds’ which fires up Deepseek instead, and burns through far less money, though I don’t think it’s as good at solving problems, just MHO.
> whenever one needs something from the other, they message each other via that file. Tasks can bounce back and forth as issues are resolved and tested.
Why isn't this just one coherent agent loop with subtools/agents as appropriate? If these tasks are related in some way, having a single context would probably make it go much better.
The freewheeling messaging part is where the token bloat is coming from. I suspect that for some of us this is actually the point. I think it's a mostly form of entertainment to do things this way. The next logical step from Factorio gameplay.
Parallel agents remind me a lot about multi core compute. It's incredibly easy to take a single core product and make it run much worse across a lot of cores.
Why not use the native inter-session messaging in Claude Code?
Frontier labs subsidising? I thought they were running with 80%-ish operating margins which is part of why they cost 10 times open weight models.
Also isn't an open weight model also subsidised? Training isn't cheap and you are not paying for it.
> This won’t last forever
It probably will. Moore's Law is still churning away in the background.
Frontier models might get more expensive, but that's a moving target. For any particular capability point, the models will only get cheaper.
Call me crazy but:
VRAM & Memory Requirements by Precision
• FP16 (Full Precision): Requires ~1,664 GB of VRAM (e.g., an 8x B300 288GB cluster).
• INT8 Quantization: Requires ~832 GB of VRAM (e.g., 8x H200 141GB).
• INT4 Quantization: Requires ~416 GB of VRAM (e.g., 8x A100 80GB)
VRAM aint cheap, Sam Altman ruined the cost of memory, Nvidia doesnt make enough consumer GPUs letting the market go insane over them, I still have friends on 1070s or 1070 TIs because GPUs have been severely overpriced for too long. I remember when a gaming PC was only $1000.
Even so why would anyone not sleep on a model they cannot run?
Seriously, if a single politician stepped forward and said "i'll bring down ram prices" they could then shoot a puppy and call me a slur and I'd still go out and doorknock for them.
Memory companies have price fixed multiple times. They've paid hundreds of millions in fines. wikipedia even has a page on it. https://en.wikipedia.org/wiki/DRAM_industry_price_fixing.
Look at the financials of these companies, they're all making obscene margins and do they plan to increase production? No. Micron is doing a stock buy back to pump the price of their share.
The Micron CEO just recently said this is the exact plan https://www.theregister.com/systems/2026/10/01/ram-supply-se...
There's sanctions, tarrifs, and a DOJ who doesn't give a shit. Until we can fix that the insanity will continue. Phones will be unaffordable. Laptops will be obscene. Gaming consoles will be thousands of dollars. Desktops will be dead.
If you're waiting for some David Ricardo equation to happen, tough cookies, it's not coming.
The market is legally locked down and we're in hostage pricing mode.
And what's the story? You can't afford electronics because we're using it to build robots to take your job? I mean ...
Nobody is coming to save us. That's our job.
> "i'll bring down ram prices"
wonder what voting would be like?
gamer vote ++
datacenter hater vote --
datacenter lobby ++
micron lobby --
Wouldn’t cheaper memory make it easier to bring compute out of data centers and onto consumer hardware?
Datacenter haters will read this as "that's still evil AI", and everyone else hopefully can count and understands it'll be worse for environment.
Yep, that's all the voting blocs.
> do they plan to increase production? No.
Micron has 3 brand new fabs currently under construction, 2 Boise, 1 in New York as the first of 4 planned for a campus.
Plus expanding other existing facilities.
These things take ~3-5 years from breaking ground to full production. You'd have had to anticipate the current demand years before it happened in order to be bringing production on-line before 2030 or so.
Samsung and HK Hynix also have fabs under construction and planned.
CXMT started 11 years ago and only now is reaching any real volume. If they decided a year ago to react to the current demand cycle they'd be 6-7 years out.
Not much you can really do to wish for more fabrication to exist on any timeline not measured in fractional decades.
Could they do more and react quicker? Probably, but everything I've read on the subject seems to point to 3 years is absolute bare minimum if you happen to have a shovel ready project with the land bought, local permitting completed, infrastructure extended to the site, and a skilled workforce already in place. They could suspend buy-backs/dividends today and dump it all into building production and there would be no material impact until around 2030.
> The Micron CEO just recently said this is the exact plan
CEO simply stated the demand pressure will not go away through 2027, and supply will not increase until around 2028 when currently under construction fabs start shipping volume. The article does not support your statement.
Stanford tracks RAM prices in this nice little site: https://dam.stanford.edu/memory-prices.html
Costs did go nuts, but there are signs of easing in the market of late. CXMT is starting to have an impact and priced will probably fall in 2027.
https://pcpartpicker.com/trends/price/memory/ is better. pcpartpicker has the data.
Yeah, but why would they make consumer memory when HBM for GPUs is much more profitable?
Capitalism eats itself this way. Second and third order effects will collapse the demand.
You need to keep the market healthy, not some insane Bitcoin style HODL pump - that's how you get wrecked.
I mean I'm not a neoclassicalist but I've read all of them. I'm in consensus with them here. There's a bunch of theories on what a healthy market is but what we're currently seeing matches none of them.
It's short term profitable but long term disastrous, especially in a world where new mathematics and techniques could literally collapse the demand overnight.
Imagine if some paper hits arxiv and the 256 GB requirement for some model now becomes 64. Woops!
Some clever trick about how attention heads and context Windows work could potentially slash a bunch of requirements by giant margins and all they're doing is firing the starting gun at that global race with every obscenely priced unit they sell.
But if prices were reasonable, this wouldn't be an apocalypse. It'd be fine. Consumers wouldn't rush to 64GB, they'd say " Cool I can multitask now at 256" or " great I can do horizontal scalability' or something else.
But no they created the market conditions so now what would happen is the consumer will immediately flip the 192GB they don't need on eBay, hoping to snatch a profit before the prices tank and the second hand market will be flooded the rug will be pulled out from the luxury pricing and everyone will get screwed.
This has happened in electronics markets before. Many times.
When Engels talked about the grave diggers of capitalism they were looking at it through a 19th century labor/manufacturing lens but arguably this same dynamic is at play here.
What "second- and third-order effects" do you suppose will collapse the demand for RAM? The people complaining most loudly about RAM costs are the people who want to run local models; if that becomes popular it will supercharge RAM demand, because locally-hosted models can't parallelize runs from many users the way cloud-hosted ones can. I don't see any slackening in RAM demand at any point in the foreseeable future, even if the big AI companies all go bust.
This is all hypothetical and debating hypotheticals isn't productive so let's roll back to markets.
Let's say ram used to cost $100 and now that same unit costs $1000. You paid say $500x1,000 for that unit during the price increase or some price where you can currently flip for profit.
You have a very expensive data center and you're in debt financed on the premise that you have these special computers.
Now a new technique comes out and it turns out you only need 1 memory unit for something that used to require 8 or 4 or some meaningful multiplier.
This stuff happens all the time. It's why we don't use BMP files on websites or serve giant MOV files on YouTube. It's why postgres queries are faster now than they were 10 and 20 years ago.
You rent out your machines. You need to service your debt.. Demand may 8x overnight to accommodate but you have a monthly bill to pay and that's unlikely. It's likely going to drop.
Think about it. Your customers are paying maybe $10,000 a month and serving their customers. Now they can drop that to $1,250.
On market if you were to sell some of that ram you have 100% profit right now but not for long.
Jevons paradox assumes unlimited capitalization, zero debt servicing, infinite time horizons...
We live in the real world so what do you do?
Historically the answer has been "sell that shit"
There's an aphorism for this "stairs on the way up elevator on the way down"
If we had a healthy market with sane prices where you can't flip the thing you bought for 100% profit the answer would be "create more value."
>Your customers are paying maybe $10,000 a month and serving their customers. Now they can drop that to $1,250.
Or they could stay at $10,000 per month since they are willing to pay that much already.m, so they just use AI more and in more places.
> locally-hosted models can't parallelize runs from many users the way cloud-hosted ones can
Why not? Unlike many other workloads, LLM inference actually seems pretty suitable for decentralization (effectively stateless means no availability concerns; bandwidth and latency are relatively forgiving too).
I think locally-hosted models at the org level will definitely be somewhat popular, but you seem to be talking about decentralizing for people's personal, non-business use, and I just don't think that's going to happen to any real degree.
People who say they want local runs really mean it: they want local runs on hardware in their room, not on some decentralized system which, if it existed, would almost certainly just be a worse, less-reliable version of cloud hosting. I'm not saying nobody would use it, but it sounds a lot like things like IPFS, which have also completely failed to displace either cloud storage or buying a bunch of disks for your own private use.
Some people will care a lot about keeping their data on-prem, but many others probably won't, and the former can then resell their spare capacity to the latter.
Decentralized storage is much harder, since there reliability matters a lot more as it's inherently stateful. You have to assume data loss, so you have to replicate everything; with inference, you only have to spend extra resources at failover time. Also storage can't be time-shared in the same way as compute; if it's full, it's full even when not actively accessed.
> The people complaining most loudly about RAM costs are the people who want to run local models
This is a tiny percentage of the population.
Samsung is cutting phone production because of RAM prices.[1] The consumer market is badly affected: budget phones, laptops, general electronics.
The budget segment of sub $100 devices in India has been almost wiped out. Manufacturers cannot afford to spend 50% BOM on RAM+storage. Unless employees are getting a 15-20% wage rise this year, I expect a similar situation in most places.
Between the engineered conflict in the ME triggering O&G price rises, and stratospheric RAM pricing, the situation is pretty bad.
[1] "There is no profit even if we sell"…Samsung to cut smartphone production by 30% (https://www.mt.co.kr/en/tech/2026/10/08/2026100709554237233)
> Imagine if some paper hits arxiv and the 256 GB requirement for some model now becomes 64. Woops!
If you were a DRAM manufacturer, isn't this exactly the kind of thing that would make you think twice about investing years and $billions in new fab construction?
The second Micron boise fab hasn't even broken ground yet, they are still working on the first one. So don't expect these things to be completed in parallel.
Some of my family is pretty happy, though, with the job security as they are pretty convinced these projects are all going to take much longer than what's being stated publicly. Micron is saying the first chip from the new fab will be in 2027... though they also predicted it'd be 2026. The date seems pretty slippy.
Anyone who has been around the semiconductor industry since the last century will remember various huge fabs e.g. in Arizona that were partially built but never finished due to oversupply by the time the walls and roof were done.
If memory prices cool, in about 2 years, Micron will stop new projects, they have done it before.
Especially given CXMT has been able to scale up much faster than what most people expected, only reason their isn't a bigger impact is modern HBM is hard to CXMT even today.
We are likely to see supply double in the next 3 years, but demand even out with optimizations, cooling of data center demand, and most importantly moving some of the dram to flash demand instead which is much easier to produce and scale.
> CXMT started 11 years ago and only now is reaching any real volume. If they decided a year ago to react to the current demand cycle they'd be 6-7 years out.
It's taken them this long to catch up to the DDR5 standard. They've only recently been through qualifications to be a DDR5 supplier for the big boys.
> Every Major Motherboard Maker Now Validates CXMT DDR5
https://www.techtimes.com/articles/321572/20260725/every-maj...
After their recent IPO, they have more than enough cash to ramp up in a major way.
It's just a matter of time.
Neither political party cares at all about memory pieces get real lol
wait until holiday shopping...it affects the price of almost everything with a battery or power cord.
What do you think they'll do? Neither repubs nor dems will touch ai companies in a meaningful way. Anything China does wrt memory fabs week be more significant
I don't have faith in the political parties. Everything is insane. You look at platter recently? It's up 3x in 12 months, not just ssd or nvme, but straight up traditional platter.
https://web.archive.org/web/20250612003557/https://diskprice...
https://diskprices.com/
After 70 years of decreasing computer prices all of a sudden it's gone 3x, 5x, 10x up in 1 year, we are in total clown world and saying "dur AI" is lazy and doesn't map to reality.
It's Argentina style inflation - as if Honda said "we're only making $500,000 luxury cars now. Everything under $50k we've stopped." and then those cars shoot up to $125k.
It's destroys the market, destroys the consumer, destroys the company, dismantles everything, and they do it for the short term payday.
I don't really understand why. Memory is a critical component of every computational device.
That's tautological, but I think you would need to explain how it extends their and their backers' influence to get party attention.
RAM manufacturers are bidding against NVIDIA and everyone else for the same constrained supply of EUV machines. And it takes years to build more fabs. Micron has multiple fabs coming online in 2027 and 2028.
Don't believe Nvidia has any fabs of its own.
> they could then shoot a puppy and call me a slur
I know it's just a figure of speech, but damn. I laughed out aloud in public just reading this.
Kristi Noem would fit the bill, except for the bit about lowering memory prices.
If we take some time to understand how HBM memory is manufactured (with particular focus on yield risk for final packaging steps), we will hopefully learn that the current capacity crisis is not bullshit.
I guarantee Micron & friends are not intentionally orchestrating their business such that they would suffer a massively reduced chance of yielding on a per-die basis. Unless someone is actually buying HBM devices, they are not going to be making them. These are not a commodity that can be speculatively manufactured in any economically rational way.
How many people would you allow them to kill?
>Seriously, if a single politician stepped forward and said "i'll bring down ram prices" they could then shoot a puppy and call me a slur and I'd still go out and doorknock for them.
How many people, outside of tech geeks and megacorps care about RAM prices? And how gullible would you be to BELIEVE the politician they could actually make it happen, and even if they did, that it would extend to the average person, and not JUST megacorps/megadonors?
Phone companies have been differentiating their models based on RAM for a decade. As have laptop and desktop sellers. The reason your router sometimes randomly crashes could very well be a result of not enough memory. The reason it takes such a long time to launch some programs repeatedly is because you don't have enough memory to cache it. Swapped from your browser to an app on your phone, but when you go back to the browser the site has reset and you lost everything you were working on? Not enough memory. Etc.
I think a lot of people care about the downstream effects of memory prices, but I agree with you that they may not realize that they happen because of memory prices.
>The reason your router sometimes randomly crashes could very well be a result of not enough memory. The reason it takes such a long time to launch some programs repeatedly is because you don't have enough memory to cache it. Swapped from your browser to an app on your phone, but when you go back to the browser the site has reset and you lost everything you were working on? Not enough memory. Etc.
That might be true at micro level, but at the macro level more memory just means developers get more lazy with their optimizations, causing apps to get more bloated, eating up any gains in extra memory. There's no reason why slack needs 1+GB to run, yet people are perfectly happy to put up with it.
> Seriously, if a single politician stepped forward and said "i'll bring down ram prices"
Or abolished VAT (the meaning of VAT is that you pay a "rent" for all the infrastructure used to produce the thing) and import taxes (protect your market) on stuff we don't produce in our markets anyway.
The situation is actually much worse and the long term consequences will start materializing soon. The wholesale theft of humanities soul is in progress. It won't be a pretty sight in supposedly civil first world countries, when the human spirit awakens. Currently we are still pressing that snooze button hard and repeatedly, as I think most are keenly and deeply aware of what needs to happen but that too will cost our souls.
I keep arguing that memory needs more competition, and people keep pointing out that it's too slow and expensive to ramp up. But if the threshold to enter that market is so steep, that means it cannot function as a free market and requires regulation.
In this case I think investment in more production is the only option, and it needs to happen even if it is expensive and slow.
> they could then shoot a puppy and call me a slur and I'd still go out and doorknock for them
Priorities
Yeah, I was about to say, the memory industry has been found guilty of price fixing multiple times.
> Seriously, if a single politician stepped forward and said "i'll bring down ram prices"
How? Increase production? The time needed to scale up the production is longer than one election cycle.
I think, considering the size of this model, it's closer to a Pro than a Flash on everything other than speed
He doesn't ruin the cost of memory. Advances in memory size and speed are now in full speed mode. Expect drastic increase in the upcoming years. Big factories are in the making and planned. Gigalab in the US and many others in the east. Since 2010 we have computers with 16gb as being normal. Finally we are moving into a new era where the standard will be 64gb next year and 128 in 2028. Hopefully we reach 1tb in 2030.
I've read the exact opposite, vendors are reducing the standard from 16GB back to 8GB
That's only temporary till production meets demand again.
Which is not going to happen in the upcoming years
one can make a fine gaming pc for ~$350
1660 ti, 4790k, 16gb ddr3
There's no BF16, original full quality weights are quantized already and 510GB.
Then good portion of those weights are n-grams (~200GB) that don't need to be in VRAM.
Then KV cache of that model is super lightweight at ~1GB per 1M tokens. If HBF succeeds, then accelerator with 16GB of VRAM and 1TB HBF/NAND is probably all you need (?).
What are you talking about? The model is native NVFP4, why you run it at any precision higher than that?
I just un-retire my pair of 1080Ti for some small models development because the current GPU prices literally make me sad.
The bulk of its weights are natively MXFP4. And engram values don't need to be in vram.
Not quite: not all of this needs to be in VRAM
It has a set of n-gram tables which you can stream from system RAM or even NVMe
That said it’s still quite big! I can’t fit it on my DGX Spark, though I believe you can if you have two?
I have access to two and will explore this the coming weeks.
Also give GLM 5.3 Flash a try: it’s shockingly good too in my testing, and I believe eugr has a TP=2 recipe to use for sparkrun
It needs 3-4 Sparks to run well (at an acceptable quantization and sufficient KV cache):
https://github.com/christopherowen/spark-ds41f
Ah that’s a shame. GLM 5.3 Flash is honestly as good IMO and can run on two pretty successfully from what I understand.
I’m quite spoiled with how good Qwen 3.8 Flash Next is on a single spark though: shocking how good local models are getting on attainable-ish hardware
DeepSeek V4 Flash runs well on two Sparks, I documented that here:
https://blog.jonathanpage.com/
GLM 5.3 Flash runs fine on two Sparks and Qwen 3.8 Flash Next on one is indeed incredible! I made this 3D game with it in two days using Qwen Code as agent:
https://games.jonathanpage.com/
I forgot to say, DeepSeek V4.1 Flash runs great on two DGX Station GB300s!
https://www.storagereview.com/review/dgx-station-gb300-clust...
Having Flash Next local at 150 t/s with 250k context is a joy. It’s as good as Sonnet 5. It will spaz out but it was less eager compared to DS Flash 4.1. Both are good but I find I prefer Flash Next. This and Qwen 27B are the models people should be freaking out about.
It's dangerous to go alone. Take this: [1]
I reimplemented most of the features of the Deepseek v4.1 flash paper (apart from quantization aware training which doesn't make sense because my implementation uses float32 precision anyways)
I'm currently learning how to distill reasoning traces (check my other github repositories) but I think that a locally selfhostable deepseek is possible with my mixture of experts sharding mechanism. I decided to optimize everything for CPU parallelization, with the idea that the KV cache and meta model have to run from CPU RAM anyways, so the experts can also be loaded/unloaded at runtime if needbe, to save more RAM.
My assumption is that the KV cache optimizations in combination with the CED and compressed attention features are the reason why v4.1 flash has so few hallucination problems and such a strong self-lookup/thinking behavior. But that's more a gut feeling, need to evaluate and test this more thoroughly.
Anyways, would love to see someone train this on their own datasets. Currently my pipeline is kinda optimized for parquet and zim files.
[1] https://github.com/cookiengineer/gonano
> Even so why would anyone not sleep on a model they cannot run?
Because it's an open model so providers compete on price.
The article isn't just about running locally though. The author is saying it's super cheap to run the model through Opencode Go (and presumably OpenRouter etc.) Personally I'm always most excited by models I can actually run locally, but even these huge open source models open up the competitive landscape for companies to let you call models via an API or just lease compute. And they don't have to charge you to offset research, training, huge staffs of the best minds in the world, crazy PR etc. I think that's a big win for customers and buts competitive pressure on the frontier labs as well.
You can run it locally for the price of a decent car, or run it (hopefully) privately on somebody else's hardware at vast.ai or a similar provider for much less. What's not to like?
No, you won't get frontier-level intelligence on a 1070Ti. Yes, it should be illegal to do what Altman did. Since we clearly don't live in the best of all possible worlds, we need to settle, and DS4.1 Flash is a good place to do that.
For tasks that don't require vision I personally like the NVFP4 quant of GLM 5.3 from Local Inference Lab better than DS4.1F, but they are both well beyond awesome.
I did somet math and completely gave up on the idea of trying any worthwhile local model and figured I'd rather pay the 15-30 USD per month via subscription and/or API key combos for years than buying a local setup which might go out of date very fast, if it doesn't goes kaput just out of warranty. I won't be surprised if RAM scarcity is an concerted effort to herd people towards the remote models :)
Thanks for the data!
Allow a question from someone who’s only got a very vague idea of how this kind of stuff works behind the scenes: say I rent usage of this model through one of the many LLM hosting providers out there, and let‘s assume I use it extensively through something like Pi or OpenCode and vibe code away all the time, keeping the hosted model occupied as much as I can, happily burning my credits.
Does that mean that there is a hardware cluster as described by you above that is crunching away just for me?
So at FP16, I alone keep a 1,664 GiB system occupied all the time?
No, a cluster can server multiple users at the same time, providers cap the tok/s so that one cluster can run inference on multiple inputs at the same time. OpenAI with their new ultrafast mode is probably reserving the whole cluster or prioritizing requests of ultrafast users above others with a higher tok/s hence the high price and high speed. There's many other knobs providers tweak that they don't show the users, for example I doubt many providers are hosting the full FP16 version.
It's not based on rate limiting at all.
The "expensive part" of generating the next token is streaming in the model weights from memory. The computations are relatively simple, which is called a "low arithmetic intensity" in industry jargon.
So what they do is batch multiple chats together and compute the neuron activations for all of them together.
This is vaguely similar to how some database engines work, where if multiple users need to run a "whole table scan" query, the additional users "join" the streaming workload of the first query mid-way, then loop back around to complete the first part that they missed. The AI accelerators don't do this looping, but the concept is the same: amortize the expensive I/O over multiple computations running in parallel.
The "turbo mode" token rate thing is almost certainly your query getting sent to slower or faster hardware, like B200 vs newer B300 kit.
It depends hugely on what "rent usage of this model through one of the many LLM hosting providers" means. If you're asking them to host the model privately then yes, all of that 1.6T of RAM is likely in use holding weights, activations and KV cache by an inference engine that's only getting/answering requests from you alone. When you aren't actively using the model the hosting process is still active and waiting with all of that memory still wired to it.
As background: For the most part VRAM oversubscription/paging/swapping isn't a thing in the same way that RAM for a VM often is. There are some approaches to it, but (to my knowledge) not at that sort of scale.
There are some systemic reasons for this, but very broadly speaking the GPU vendors are building toward the highest bandwidth and lowest latency possible, and the overhead/complexity of something like protected memory modes serves neither of those priorities.
This “blame sama for memory prices” meme is so tired.
He gave demand signal so many times years ago and was mocked for it and now we have the consequences of industry not taking him seriously.
Projects like DwarfStar https://github.com/antirez/ds4 really lower the hardware bar a lot so Deepseek 4.1 flash and other mixture of expert models can run on consumer hardware. There are also other inference providers who make their money serving openweight models. Services like OpenRouter make it all too easy to utilize these models. Access to these models isn't hard. The hardware moat is becoming pretty easy to bridge.
More concretely DwarfStar M5 128GB Deepseek 4.1 flash 1K tokens @ 29s, 5K tokens + reasoning @ 147s, 10k token prompt @ 463 tokens/s = 22s. Hardware buy-in USD$7K / AUD$8.5K / EUR€6.8K. At typical workloads, ROI is still poor vs. current-era subsidies, but owning hardware is good for privacy/longevity/connectivity independence. Whether you actually consider Apple hardware 'owned' is a valid and thought provoking question.
Still gonna take 2-3 years to get DeepSeek V4.1 Flash quality at decent speeds on reasonably priced hardware.
Hardware update cycles are 2-3 years even on the high end, so it's still a ways away before "good enough" and "local" belong in the same sentence for the average person.
And by then, DeepSeek V6 Flash will be too cheap to meter, 5x faster, and 10x better, so... You'd still need to go out of your way.
Most people are spending most of their time on their phones anyway. ..
In the words of a Scottish comedian, "The average person is Chinese." https://youtu.be/LsEhNMy8svo?t=98
Flash Next is a basically there. It really depends on what you are doing. This model is great. People forget that they felt Opus 4.6 was a great model and now you have it at home.
"1070s or 1070 TIs because GPUs have been severely overpriced for too long" ... ."
1070ti launch MSRP was $450 ish. 5070 could be had in the last year for 5xx-6xx range easily.
All things considered - (inflation being about 30%~ (guess)) between these two timelines. You are looking at 300% performance difference at a cost dollar for dollar that is cheaper then when they purchased their cards.
Might be a bit of a stretch blaming it on "severely overpriced for too long..."
The 5070 honestly feel more like xx60ti class instead of xx70 class
Are you counting the n-gram/PLE as part of the model weights there? They can go in host memory. Would be good to show your working. Also the released weights are pre-quantised and presumably QATed, so your "Full Precision" and INT8 are simply not a version of the model that actually exists.
Edit: I went and checked for you. The LM backbone is 307.2 GB (286.1 GiB), straight from DeepSeek's upload. The n-gram table is 203.1 GB (189.1 GiB), which goes in host RAM. Note the embeddings are higher precision than the expert tensors, so it's a larger fraction of the bytes than it is of the parameters.
So,
> Call me crazy but:
You're crazy. :-)
DeepSeek V4.1 Flash is mixed MXFP4/MXFP8 so all but the INT4 calculation is wrong here and that's still wrong because you can run it on < 400GB VRAM. The n-gram table is MXFP8, but can be offloaded to RAM or disk without too much of a performance hit. Really, you could probably cram it onto < 300GB VRAM if you're willing to apply a small quant to certain parts of the model considering how little VRAM is dedicated to kv cache.
I know my comment is a little nit picky because it's still pretty expensive to run, but it's not quite as bad as this comment makes it out to be. Really, if you're VRAM constrained, take a look at GLM 5.3 Flash or Qwen 3.8 Flash Next before you worry about this model as all three models perform pretty similarly.
In nvfp4, it's about 300 gigs once you offload n-grams, 491 without offloading, you can run it pretty well on 4x DGX Sparks, which last I checked was about $20k. So, it's definitely runnable.
Or you can just use any of the neoclouds' shared hosting. The thing for them to be freaked out is that these models are getting good enough very quickly, and all the shared hosting providers can run them for a tiny fraction of what the frontier model companies charge.
what they should freak out about is qwen 3.8 github.com/Niko1221/Strata
it rips with just 64 ram and a 9070xt
I'm paying for the heavily-discounted subscriptions, not the API rates. There isn't really a cost gap for me. DeepSeek doesn't have a subscription to compare to, but when I compared the GLM 5.3 usage I got from a $100/mo Z.ai subscription compared to Opus 5.5 on a $100/mo Claude subscription, there wasn't a big gap. And GLM 5.3 is very clearly not a frontier model (deepseek v4 seemed a lot closer, but I didn't use it enough to really say for my workloads).
I don't think those subscriptions nave negative contribution margins, either. I think we're seeing a lot of price discrimination by the big labs, and huge margins on their frontier models. The fact that they have been cutting prices to their second-biggest tier of models (Opus/Sol).
Open models catching up and collapsing these margins would worry me if I were a shareholder in the big labs, but as a user, I really doubt that the western labs have bigger environmental impact just because they have higher API costs, I think they have a ton of efficiencies they aren't sharing with customers yet because demand is so high.
The tightening of subscription value has already begun. dsv4.1f is already worth paying for at market API prices. Maybe it goes to 2x because apparently no one has figured out how to match DeepSeek's insane caching efficiency, but I don't see it getting much worse than that.
Plus you can also get dsv4.1f subsidized. OpenCode Go gives 4x if I understand their pricing correctly. Anecdotally, I feel like I get way more out of my $10/mo OpenCode Go sub for the price than my $20/mo ChatGPT, even using gpt-6.1-sol high which is very cheap, and I have yet to convince myself dsv4.1f is a worse model.
It's really not cheaper than frontier subscriptions. It's getting closer, and it's a great model, but it is not more value per task than the frontier subscriptions. Don't be swayed by the token costs, it's very chatty, like 3x more tokens for the same task as sol. I used dsf 4.1 full time for about a week.
It blows frontier API pricing out of the water, but again, look at cost per task, not token usage. Still easily wins though for my work.
I do think it's the most viable alternative I've seen so far, and that applies pressure to the frontier models. Should subscription prices hike or become unavailable for some reason, I know what I'll be using.
When pricing this, it's important to consider whether or not you want to opt out of data training. You won't get the advertised rate. Also the dsf 4.1 subscription providers are throttled af... and of course they are, because otherwise they'd be haemorrhaging money.
I am looking at cost per task (and speed per task), both benchmarks and anecdotal experience.
What are you talking about?
DS4.1 Flash not really cheaper than frontier models???
It is insanely cheaper.
Not by cost per task, just cost per token. But they're spending way more tokens per task, so it doesn't work in their facor.
I still use them because they aren't as squeamish about random things American CEOs don't like like decompilation.
For coding use cases, it really isn't.
DS4.1 flash is $0.30 in / $1.20 out (per M, peak, cache miss) Opus 5.5 is $4.00 in / $20 out (per M, cache miss)
However, that is API prices.
Anthropic offers a $200/mo subscription. How this translates into usage is admittedly a bit opaque, subject to change, and depends on how exactly you use it. But it's a lot of usage - Semianalysis data shows that $200 is getting you around $2,500 of usage at API rates if you use Opus 5.5. This is close to what I'm seeing anecdotally with my accounts, if anything I have been getting a bit more.
Now, unlike DeepSeek, you can't use your subscriptions to power live AI-driven products, or resell tokens in any way. But for personal coding agents, you can use as many of these subscriptions as you want, for now. So I am paying effectively basically 8% of the published API rates, so at my usage:
DSv4.1: $0.30 in/ $1.20 out Opus 5.5: $0.32 in / $1.60 out
Obviously, those aren't real prices, but they accurately convey apples to apples what my everyday usage costs me and most other heavy users, and why it's so easy for me to stick with Anthropic/OpenAI.
I don't think it's a coincidence, either - I think the token allowances for these subscriptions are set to be competitive with the open models, so that most coding users (and their incredibly valuable data) stay with the frontier labs, while VC-funded wrapper companies and less-price-sensitive giant companies with strict procurement policies pay exorbitant markups for enterprise contracts at the API rate.
Locked into their tools though. I happen to be very tied to a IDE centric model (old dog can't learn new tricks) and their Desktop thing is a regression for me. I can use the cli, but it wrecks the muscle memory I have with what I use now.
And Anthropic is somewhat unusual in that 10x more tokens via the subsidized path. I imagine more rugpulls are coming.
Not disputing your point in any way, just noting there's already caveats, and more are likely coming.
> It's really not cheaper than frontier subscriptions.
It depends on how you use it. I used to have the $100/mo Claude plan. I would easily blow through limits when I was on the $20/mo plan, but would rarely hit them when on the $100/mo plan.
Lately I've been using GLM 5.3 Flash (from Fireworks), and my spend is $1-$2 per day when I use it for coding, so max $60/mo (less, since I don't use it every day). IIRC DeepSeek 4.1 Flash is priced similarly.
If I had to pay API rates for frontier models, I can't see how $2/day would cut it. Maybe GLM/DS are chattier, but not anywhere near the 10x required to make the price difference not matter.
Sure, if you're running agentic loops all day, 5 days a week, you're probably going to blow past even $200/mo in API charges pretty quickly.
IME GLM is inferior compared to Deepseek v4.1 Flash, highly recommend running it through some real work
Marketing.
Have you guys see how aggressive is the push for enterprise use by both OpenAI and Anthropic? I had friend from a non-tech industry in Asia telling me that their company was offered free trial of the enterprise version of Claude, with trainings and such.
On the other hand, DS and Z.ai, have zero to none marketing outside China. There is friction to use DS/GlM models and the ZDR is unclear, so most enterprise that has heavy AI usage hasn't move over yet. They would rather spent $200 for the peace of mind than to take the risk of being slam as a national traitor down the road (which again is another form of marketing by Big AI, trying to frame Chinese models as thiefs).
So, I don't think they are not freaking out, it's just that they are addressing different market segments and reacting to the situation differently.
Aggressive marketing (including daily posts on HN). Many people simply don't know about alternatives, there are many people who never heard about pi and opencode and live comfortably in claude-codex bubbles.
I have been using DeepSeek 4.1 flash intensively for over a month. If I run it all day long it costs $1-2. Its fast. Previously I was always quickly running up to my Claude/Codex 5 hour window (on the $20/month plan). The cost savings of DeepSeek is real as shown in this article and I am using subsidized plans.
DeepSeek is horrible at grilling sessions (the /grill* skills to make technical decisions). It doesn't know how to explain things. Maybe the skill could be adjusted. It also doesn't come up with as good solutions as Opus/Sol.
What I use it for is
Previously I planned with Opus/Sol/Astra and then I used DeepSeek for coding, and then reviewed with Opus/Sol/Astra. With the cost improvements to Opus/Sol I am trying to use them for coding instead now so there will be less back and forth review needed.
They are all working together in Pi using the extension @tintinweb/pi-subagents where my workflow skill is calling different subagents that use different models.
Luna is cost competitive, but doesn't score as well on intelligence. I do need the intelligence for most of what I use it for, so I am not motivated to use Luna. Haiku also doesn't seem like a competitive price/performance mix.
Luna is 1 point being on AA's index at 1/4 the cost, yes it "doesn't score as well" but paying 4x for 1 point is crazy if you're going off benchmarks.
AA has Haiku 5.5 as cheaper than 4.1 Flash (both on Max, which isn't ideal but what can ya do) and a 4 point intelligence gap.
Why do people like to think open models are more competitive than they are?
DeepSeek's own paper advises against using Max, showing that it normally doesn't perform that much better. I am not using it on Max, so that's not a useful benchmark for me. I have seen other benchmarks where Flash does significantly (30%) better than Luna.
It is super bad on a bit more complex workflows and starts repeating same errors with the same tool until the cycle breaker hits.
6 is worse than 5.6 here.
But it is amazing on generating a report on content generated by better agentic models such as DeepSeek or GLM, which both do a mediocre/bad job on reports.
I'm using DSv4.1 in OpenChamber (eg OpenCode) using the Superpowers skills and a lot of custom AGENTS.md instructions to iron out the kinks and I genuinely cannot see a difference between it and Opus and I've been building native iOS and AppleTV apps, Go servers, Typescript, Cloudflare workers, Svelte/Astro, etc.
It's a super capable model all around from my experience.
How are you getting down to $1-$2 per day running "all day long"? I've been using GLM 5.3 Flash and I also spend $1-$2 per day, but my use is pretty modest, I think. DS 4.1 Flash is priced similarly to GLM 5.3 Flash; can't imagine DS is significantly more token-efficient.
You can't use a flash version and complain it doesn't think well you would use the pro version.
While I agree, this particular discussion chain really frustrates me.
1. "This Flash model is really smart. Here is an article to discuss how smart it is. Why aren't people freaking out about how smart this Flash model is?"
2. "I tried using it for a smart thing. It doesn't work so well for it."
3. "You should know better than to use Flash for smart things. It's not meant for smart things."
Because DeepSeek is not "a month or two" behind as claimed in the article.
These open models still did not beat February's Mythos / Fable 5.
DeepSeek 4.1 Flash is behind GPT 5.6 Sol, and that one is left in the dust by the excellent Opus 5.5.
Rumors say Anthropic is holding in reserve the big improvement, Fable 5.5, for the IPO.
It's plausible that open models are 6 - 12 months behind, and there is no "good enough". As long as progress doesn't slow down, leading labs have nothing to fear.
I was thinking about this earlier today and I came to the following question:
If you had a model 10x as capable as the best model out today, but it cost 100x more, would there be a market, and, if so, how big?
I think there would be a market and I think it would be large.
So, I agree.
Doing what? How many jobs involve solving Millennium Prize math challenges?
99% of everything is CRUD LoB apps.
I am coding CRUD apps with a mix of astra, sol 6.1, fable and opus 5.5. A more capable model would still benefit me imo. Being able to follow high level guidance better, and being able to harness other models for each task would be a big improvement.
Do you know how what you're doing, or do you find yourself working on things you dont understand and need the best model because it's the only way to push your own capabilities (because you're avoiding learning how to do the thing yourself)?
Not asking to be mean, I just genuinely dont know why you'd need the frontier for basic applications.
It's a matter of bandwidth. The more I can offload onto the model, the more I can accomplish. For example, I had to do a lot of security work over the last 2 weeks to get ready for an event. This requires handholding current models on many fronts, like: 1) Do they actually implement the security fixes correctly. 2) Do their fixes create any new edge cases. 3) Do their fixes compromise existing interfaces or API surfaces.
I cannot trust current models to find all the necessary context, or to make what I consider to be good trade offs. A much more capable model would be able to see my existing patterns (or at least not have context rot make them blind to my convention docs) and make trade offs I agree with much more consistently, and I'd be able to do more with my time.
I've actually found models to be pretty poor at driving things I don't know well, so I generally don't do that unless its general design/product exploration and the end product code is throw-away.
Many people would, and you'll find that they're building crappy webapps where you dont need SoTA. Like seriously who needs these frontier models?
Unless you're doing some extermely difficult post-grad lvl research, you do not need a 100x PhD research assistant, especially not for whatever silly SaaS product most people are building.
There's people at my job that get so much more done than everyone else using Fable/Opus/Astra. and all they use is the fastest cheapest models. I'd say the people who are using sota models for everything are doing it just because they prefer to be lazy.
You simply do not need these frontier models, they outgrew most people's needs 6 months ago, but for some reason people still want to run a 700k rack of gpus full throttle to center a div for them.
I agree that for average web dev tasks the open models are already good enough. I've had good experiences with both DeepSeek and GLM. And these models are just better for anything security related since they don't throw massive hissy fits.
However I do actually have a project where I need the frontier models--I'm working on a deep learning project of moderate complexity (something novel/state of the art within its domain, adapting a known approach from published research in a related domain). The difference from Opus 5 -> Opus 5.5 was huge for my project. Opus 5 was struggling, Opus 5.5 is doing really well.
I think the demand for frontier models will continue to be there, at least for a subset of tasks, although I agree that it is probably going to shrink as the non-frontier becomes more and more capable.
I think the market would be huge, especially if it's the "can complete a large task in 3 turns instead of 15" kind of smart. Lots of people and companies would pay for quality + speed.
I think there would be a market, but it would mostly be a FOMO market. That is, people would be doing tasks on it that the "regular" model is more than capable of handling, because they're afraid they're leaving something on the table by not using the absolute best option.
Certainly, there's a real market for it too, with people who would actually use its advanced capabilities, and see the 100x price as worth it.
But sure, even a mostly-FOMO market is still a market. If people are paying, people are paying.
What does 10x more capable look like now? Surely at some point we will reach an asymptote of what can be done purely digitally: all useful coding tasks can be automated, most math research, etc. At some point the physical world becomes the dyke holding back the singularity; until these genius models can scale their investigation into physical experiments and manufacturing, the future will have arrived only in the digital world.
I think you're making a mistake in thinking the digital world is the only one reachable to AI. Robotics and sensing would be opened up by a sufficiently capable AI.
I agree: large market. I think the future of these frontier labs is selling exceptionally powerful and exceptionally expensive models. They'll be used for precision, high value tasks. The rest of us will be happy with good enough and cheap models.
There absolutely is "good enough" and I agree with this author: DeepSeek 4.1 Flash is plenty good enough for all the things I would trust an AI to do at my job.
Yes, trust is exactly the point. I trust the frontier models to tackle bigger chunks of work than the open weight models, and get it right.
that just sounds like openai/anthropic cope/propaganda, based on absolutely nothing objective lol
even their harnesses are far surpassed by pi and opencode at this point
also sick 'rumors' lmao, apparently marketing through rumors is in vogue these days
>> that just sounds like openai/anthropic cope/propaganda, based on absolutely nothing objective lol
Nah. There are benchmarks. They are free to look at. And they paint a very clear picture.
Yeah the picture they paint is that they're mostly bullshit
Got it. Thanks for your input I guess.
Please don't be sarcastic, it's against the site rules
I see people throw around these benchmark numbers
I've seen different benchmarks come to different conclusions
Benchmarking these models must be an incredibly complex and difficult problem
How can a lay person know which benchmarks actually have good signal?
Agree, will see after the dilution and CoT hack fixed, will they keep the pace now. MiMo had some good numbers recently because it's discovered that the post evaluation RL directly exposes answers to models, so RL and evaluation is runied.
Perhaps on certain benchmarks and for certain work, but anecdotally I've not been able to see a difference between it and Opus on a lot of dev work (web, Go, iOS/AppleTV native, scripting, general tasks)
Imo deepseek 4.1 (and a lot of the cheaper models, Luna is similar) show the issues in benchmarks. At this point.
In actual day to day development the differences are a lot harder to spot. Maybe deepseek is worse, but I asked it to run until it was able to launch itself and verify it worked as expected, and it did. Maybe it wasted some turns, idk, but when it said it was done, it was done.
I have no doubt there's things it's worse at, but what percentage of development is truly novel?
Personally I find this shocking. I don't think our application is that complicated, (typescript full stack graphql reactnative etc) but deepseek 4.1 flash is a bumbling fool, junior-level at best, who takes a very long time to make a very big mess. Opus 5.5 one shots truly impressive code in 5 minutes, while deepseek 4.1 flash takes 20 minutes to do horribly. I simply don't understand how folks claim they get good engineering out of it. Maybe we still care enough about the fundamentals to notice the mess ...
Are you using OpenRouter? I’m honestly surprised open model labs haven’t been calling them out, but heaps of providers either silently serve heavily quant versions, or don’t have inference set up correctly and don’t run the model properly.
Was a night and day difference going directly to deepseek api
In this case, I was using Hugging Face inference set to one of a few American providers. For my professional work, I do not use Chinese inference.
> leading labs have nothing to fear.
Except being priced out.
The big labs' financials are based on their products being used widely by a lot of the general public. If it turns out that they're actually selling a premium product to premium-product consumers at a premium price point (while everyone else buys DeepSeek-like cheaper/worse products), that's a big issue for them.
If a consumer computer hardware company launched by promising investors that it'd be the next Dell/HP and it turned out to be the next Apple (talking Macs here, not phones or apps/services), that'd be an issue for them too.
> These open models still did not beat February's Mythos / Fable 5.
On what task? By who? On what benchmark? How do you measure in you own workflow the “betterness” or “more goodness” of these or any models? If you don’t say those things you’re just writing a bad ad copy.
> It's plausible that open models are 6 - 12 months behind, and there is no "good enough".
Anecdotally, a lot of people - including myself - seem to really notice much difference between the model now or six months ago. So there really seems to be good enough. It depends on the task you use them for and how you measure the output. For most tasks you really do not need frontier capability. Also how do we know how much of these “big improvements” come from the harness and tooling rather than the raw capability of the model?
Did you see any person claiming an open model was more intelligent than Fable 5?
It's always some sort of "I don't notice the difference".
And honestly, if you don't see a difference between the SOTA from 6 months ago, which would be GPT 5.4, and today's Opus 5.5, you would have to be downright blind. Not sure what else to say - the results are obviously different for any kind of meaningful output.
> Also how do we know how much of these “big improvements” come from the harness and tooling rather than the raw capability of the model?
By simply running the old models in the latest harness. Which none of the people who argue "it's all the harness" ever do.
Ok so you have nothing, just vibes. Fair enough.
The benchmarks don’t mean shit. Opus 5 was a terrible model and yet it had very impressive benchmarks. All the labs are benchmaxxed to the tits, only open models’ benchmarks are even worth paying attention to because they literally cannot cheat.
Oh man. v4.1-flash has been an sbolute game changer for us. We run all our Personal Assistants now on flash (thinking high) by default and it works incredibly well. There is really no need for basic agentic tasks that might require Kimi K.3 or GLM-5.3 levels.
Once its gets juicier, we let flash launch specialized subagents with specific models. GLM-5.3 for coding or Kimi K.3 for research and critique.
But as a main driver. I love flash. And it brought our bill down by A LOT :D
Have you compared it against actual SOTA models like latest Fable or Astra?
Of course there's still a huge performance gap
but DS 4.1 Flash is good enough for most tasks
The author explains this very well tbh:
> Today's models are now good enough for high-quality unattended tasks. Chasing the latest and greatest is silly. It is fun to see the new Fable capabilities, but the tasks we throw at them are usually ridiculous (maybe even insulting) if you believe in LLM sentience. It's like asking a math PhD to organize the files on your desktop.
I'm using DS V4.1 Flash as my main model since their release and it works great for all my coding tasks. My setup is OpenCode Go subscription and obra/superpowers skill.
The only times I try to change models are on general planning tasks (like research this codebase for tech debt mitigation opportunities) or if I need deep research which would benefit from searching the web, in which I still think Gemini is still the best because of the speed and access to google search index. But these are not even 20% of my daily tasks.
>We run all our Personal Assistants now on flash
are you worried about sending all your data to third parties, especially if they're in different countries?
I have asked OP that question but I think there are providers who are not in China and they just host the model/inference.
Just use a provider hosting it in your country especially if your country has major data centers then its the same as using Anthropic or GPT of GCP Model Garden or AWS Bedrock
nobody here is talking about running frontier level intelligence locally so if you’re Chinaphobic and prefer layers of corporations siphoning your data in between you and the party there are plenty of options instead of directly to the party
Not shilling for them but Ollama cloud hosts domestically with ZDR afaik. I run 95% of my open weight inference through them. The rest goes through Opencode Go $10 plan (which is enough to run 3 hermes agents on DSF 4.1 and leave plenty of left to experiment with when new models drop).
Just a reminder that any API using a Cloudflare TLS certificate isn't ZDR.
The model engine provider might be ZDR, but the service as a whole isn't.
did not know. ty! have my updoot as thanks
Cloudflare doesn't retain request bodies by default, and if you don't trust that then you shouldn't trust the third party AI provider either. Cloudflare does cache responses, but that doesn't typically apply on API endpoints and is trivially disableable.
I use it through OpenRouter, which has ZDR enforcement.
what would be the concern here?
We use hosted providers in the EU. They have a cost markup but its worth it. I enjoy the idea that I do t talk to big tech.
I cant ofc be fully sure because i don’t own the chain end 2 end.
What is the cost of access like for DeepSeek-v4.1-flash, compared to GLM-5.3-flash via ZAI's Coding Plan? Because that's what I use; and often hit the "wait". I wouldn't mind trying a new model subscription or even API access which hits around glm-5.3-flash level weight class (which seem to be enough for me; with quite some human suprvision and nudging) but gives muuuuuuch moooore tokens for the same price.
(And what are the preferred providers?)
What personal assistants are you using?
Hi. We are building our own, but the core is openClaw. Hoever, that will change soon
I have a pretty large, complex project I've been building with heavy AI use (new language + compiler). I was following a 'strong model as orchestrator launching cheap models as implementers' pattern, but I recently trialled just using Deepseek-V4.1-Flash as the model for both layers because of the cost savings (with mimo v2.6 flash on code review agents for some decorrelation).
I was previously using GLM-5.3 as the orchestrator, after switching to DS anecdotally there was an unnacceptable quality loss, mostly around not taking all the relevant context into account when making decisions, pulling new design out of thin air without discussion too often, and being way too wordy and rambly in documentation despite prompting to avoid it. There's a lot of docs, rulings, core concepts, design philosophy to uphold and DS was just not cutting it.
However, it's perfectly capable of being the sole agent for all of my well specced implementation tasks. I've gone back to GLM as the orchestrator.
On the Claude side of things I was previously following "strong model directs weak" with Fable directing Opus/Sonnet (its choice per-task). Since Opus 5.5 came out I've just been having Opus direct Opus.
The sub-agent separation is still valuable to keep context clean for the orchestrator, but I just have no reason to use Sonnet as the grunt-work implementer because I'm finding it hard to run out of tokens with Opus 5.5 on a $200 subscription plan. It's really really good at subjective quality of work per token used.
I would probably go that route if I could use other harnesses with claude models, but I don't want to be locked in to claude code, and their API pricing (which you need to use it with other harnesses) is so much higher than subscription.
plenty of harnesses use the claude subscription
Ever since Claude banned this, I haven't found one. Which harnesses have you found that allows using Claude through the subscription plan?
oh my pi has it.
My work pays for Claude and cursor.
I have actually just dropped to using sonnet for everything, sure it does need some directing but I have yet to see a need to jump to opus.
To me it feels like sonnet/terra and composer 2.5 and grok 4.7 are actually good enough for most tasks and these companies are pushing the high models simply to make money.
Can't wait to just use the hand made Jai and laugh at everything else built with AI :)
You can just use plenty of good handmade languages now, I'm not claiming mine will ever be good or useful to anyone other than myself. I'm using AI so heavily on my project because I want to explore the PL design space without spending years on implementation for things I'll probably want to throw away and rewrite a different way once I actually play around with them properly. And because I lack the motivation to persist through the sheer volume of grunt work that language implementation needs to get to the juicy interesting parts.
If you're not using the Superpowers set of skills, give it a shot. It's been working really well for me on a variety of tasks.
I used Superpowers for awhile, but I don't feel like it gave me significantly better results than just raw-dogging it. It did however burn through my tokens significantly faster.
Maybe I'll come to miss it now that I removed it, but I certainly don't yet.
Because the free api is mispriced, opus 5.5 on subscription is 10x cheaper. Also, the use case for flash models are
For coding, I rather spend 10x more than have even 1 bug but I'm only spending 2x 3x more if you count subscription cost.
This is a killer use case for something like customer support though.
Its simple. My company is very willing and able to pay ~$200/engineer/month for the best version of these tools. My company is even willing and able to pay as much as $500/engineer/month, but does not need to at this moment.
My company is not willing to pay $50/engineer/month for a cheaper version that is nearly as good. My company is also not willing to pay any amount for a product produced by China, even if it is hosted in the United States.
It’s still cheaper to get a subscription to an agent harness with frontier model backing for most people. DeepSeek really only becomes appealing to me when I want to do something the monthly subscription harnesses don’t support (API access) or more harshly charge quota for (like running coordinated sandboxed subagents). DeepSeek is then great because of the low price and price transparency, it just can’t compete with subsidized monthly access.
I actually have a fairly simple answer to that: if it doesn't come up in the list of LLMs that Cursor supports, it effectively doesn't exist.
I'm well aware that there's nearly infinite opportunities to yak shave "perfect" OpenRouter setups and some people appear to enjoy bouncing from IDE to IDE as though change costs aren't a thing, but I discovered that I genuinely like Cursor and at least right now it's insanely subsidized by Auto clearly defaulting to whatever Grok's most powerful model is.
I dropped my $200/month subscription to $20/month and stick to Auto for all but really important Plan tasks, and I have basically zero chance of using up my monthly credits even using it 6-10 hours some days.
> I actually have a fairly simple answer to that: if it doesn't come up in the list of LLMs that Cursor supports, it effectively doesn't exist.
You make Cursor sound like one thousand times more important than it is. It's a product in deep water.
Because most of the people use it through enterprise agreements and don't pay the bill? I run it for my own use cases and its pricing plus caching capabilities are hard to beat, cents for millions of tokens. https://substack.com/@rubenafo/note/c-332218129?r=26y5kn&utm...
The problem with this low listed price per token is that, in reality, DeepSeek 4.1 Flash uses 10× more tokens than GPT-6.1 Sol for an equivalent task and delivers a lower-quality result. So there’s no real benefit to paying 10× less per token. Also, as someone else pointed out, OpenAI and Anthropic currently offer subsidized subscriptions for $100 or $200 a month that provide far more tokens than the API, so we should take advantage of them while that lasts.
That's my experience too. sol-6.1 goes straight to solution, like it has done it 100 times before. Deepseek will explore and insecurely overthink like it's an intern made CEO.
You can see this in the Artificial Analysis benchmarks - GPT 6.1 Sol on medium thinking scores 10 points higher than DeepSeek 4.1, but is actually cheaper per task, as it only outputs 15 million tokens instead of 250 million.
Though for longer sessions I think DS4.1 would still come out cheaper... it's hard to beat that 98% cache discount
I had the same experience using ds 4.1 last couple of weeks. It’s insanely good for the price. I’m doing mostly web dev it excels at everything I throw at it. The pricing is ridiculous. I canceled my gpt subscription and haven’t looked back hope the pricing stays like that. I almost never need a better model. I still keep my Claude 20$ sub for now but I feel like one more iteration and I won’t need even that anymore I hardly use it
If DS4.1 impresses you I would be really interested to see your comparison to GLM 5.3. I switched from the one to the other and even if GLM 5.3 is a bit slower I don't think I'll be going back.
GLM 5.3 is very impressive and definitely better, but it also at least 4x the price.
On that note I’ve been subbing in MiMo-2.6-pro when cost is an issue, which is super cheap and also performing really well.
IDK what happened today but I used GLM-5.3 as usual from Ollama cloud and it was so fast it generated entire documents like instantly.
The reasoning and the result document were done after less than 1 or 2 seconds.
Have Ollama suddenly bought GPU capacity?
DS4 (not 4.1) crossed my dont-care threshold and I genuinely stopped paying attention to new models. I'd love to try GLM 5.3 but I just don't see any point in spending the effort any more. I can get passable intelligence for a bargain price either direct from China or from a ZDR EU provider for a small markup. Paying 10x more will not make me 10x happier, it's unlikely to make me even 1.1x happier now I've got some intuition for the natural limits of these models.
I don't even bother checking how much I spent on API any more, its well under $30 over the past 2 months despite daily constant use. Who even needs a subscription at these numbers?
I’m sure it’s more intelligent but the speed of DS is so freaking fast I can’t user slower models anymore
But it also frequently uses massively more tokens for the same task compared to other models which kind of negates the speed.
Wallclock time matters and GLM 5.3, even when it is considerably slower (~30 tokens right now vs 100+ on DeepSeek 4) it is quite frequently faster on the same set of tasks overall. Deepseek 4 seems to do the 'Oh, wait' thing just about forever and has a tendency to find irrelevant rabbit holes that it then spends a massive amount of tokens on.
There is also GLM 5.3 Flash
What inference provider are you using?
The question seems rhetorical but I think there are two reasons in some combination. First is it there is some awareness lag here. That lag can be on the producer and consumer side. Software enterprises are pretty slow to adopt new things and slow to try new things so they might only be aware of openai and Claude as options. Plus there are some scariness because deep seek is a Chinese model and therefore export restricted - never mind that there are American in European providers.
The other reason is more interesting. Maybe the frontier providers think that price performance is irrelevant in light of very powerful frontier models that can start the RSI loop and or a huge displacement of work and a winner take all economic situation. After all if frontier providers earn everyone's money then you won't have any money to spend on any model 100x cheaper or not.
I think it also has a bit to do with the AI sector of tech still moving at lightning speed.
Theres already models that outdo DS 4.1 flash in cost/performance. Luna 6 on max effort for example. Luna also doesn't care what time of the day it is for cost calculation.
And I'm sure by the time people ask why Luna 6 is being slept on there will be another cost/performance king
Luna is very slow and bad at agentic tasks. DS runs circles around it and there are US providers providing cheaper rates no matter the time of the day.
While using DeepSeek v4.1 Flash I was architecting a system and I made a mistake of drawing the RPC boundaries at a wrong place that did cost me in so many ways.
I realized that mistake and guided DeepSeek where it should be.
Next I fired Fabble 5.5 set to high to check if the hype is real about Fabble. It exhausted 89% of quota and came up with NOTHING that DeepSeek hadn't flagged itself already in its notes.
Do you mean Fable 5.1? Or Opus 5.5? I'm not sure what you're working on but for me DS 4.1 flash isn't nearly at their level. For the price it's obvious very impressive, though Luna 6.0 is excellent too.
The problem with benchmarks and proprietary models is that one day a model is best at doing X, another day that's not so sure. And anyway, we are not throwing the same X.
I've found supposedly smaller and, less performant models do better on certain tasks. I end up using several models, sticking to what my unconscious statistical observations tell me to use for the kind of task at hand.
Fabble 5.1.
care to share what exactly are you working on?
I am 4.1 maxxing on commandcode GOAT Plan + api rates with oh my pi for the last 4 weeks, it's absolutely amazing and crazy fast, it's alright if it makes a mistake, I have enough time to iterate again, I have also added an advisor layer of mimo 2.6 pro which does make it a notch smarter. Getting haiku 5.5/sonnet5.5 to work on plans and letting 4.1 flash work through it is helping a ton too.
I am a big ChatGPT fan, all our team has ChatGPT Subs, but the TPS across all models including luna is just so damn slow.
Commandcode giving 60$ worth of Deepseek for 10$ is just genuinely goat.
And it never says no for cyber tasks so that's a big win
Yeah. I've been enjoying Coralbricks 250-350 tok/s speeds and it is hard to go back to slower models.
Lithos promises even faster speeds if you want to pay more.
Comparing it with Anthropic, ANY model is cheaper and more effective. Don't get me wrong, Anthropic models are good, but they're always more expensive for the same task, even compared to other closed source models. At this point, I honestly think Anthropic has played the nasty trick to fine tune the models to be too verbose and charge us for more tokens.
Not sure why the author thinks Anthropic's models spend more power and water than DeepSeek, there's no evidence of that. Their pricing has more to do with premium perception and less to do with COGS.
Everyone does optimization of model serving because it's good for every player in there.
(Also the water consumption thing is not a real issue.)
Because it’s not even that cheap? The author chose to only include Claude in their chart and ignored the fact that 6.1-sol and even more so luna can easily beat Deepseek on cost. Of course almost free cache used to be the main differentiator, raw token cost is deceptive since 4.1 just uses way more tokens than most other models
This model finally got me off my Claude Max subscription. I’ve found it superior to Opus 5.5 in certain use cases, and certainly faster.
I’m convinced that I’ll have good enough inference on my laptop at reasonable speeds within the next year.
> Sure, they stole Claude's training, and Anthropic stole it from other people. I'm not getting into the whole who-owns-whose-data debate, because most developers aren't thinking like that. They're just trying to get the most bang for their buck.
Until they get laid off and suddenly discover their moral compass.
> With my OpenCode Go sub of $10/month, DeepSeek is basically unlimited
My OpenCode Go monthly window was scheduled to reset this morning. It was sitting at 22% used despite me using DeepSeek V4.1 Flash heavily as my implementation agent the past couple weeks (I use gpt-6.1-sol high for planning/orchestration).
I had 1.5 hours left so I fired up first 10, then 20, and finally 50 concurrent subagents all working on reverse engineering C code from an old PC game. They found over 100 new functions.
This is the first workload I've found that could make a dent in my sub. It got my 5 hour window to 85% used, but sadly my monthly was still only at about 35% when it reset. So that cost maybe $2.
How are you guys doing orchestration? I have been fumbling around in my free time trying to build something for myself but is there a repo or something that just works?
I might not be the best person to ask. I use Pi harness in tmux and just ask my current agent to spawn interactive pi instances in new tmux windows, create a sentinel file for each of them, and monitor the sentinel files for signals every 2 seconds.
Currently have auto compaction turned off. When the orchestrator's context is getting close to full, I have it write a handoff markdown file and point a fresh agent at it.
I do feel like I'm getting close to the point where I might be ready for something more sophisticated, especially wrt to subagents communicating with the orchestrator.
That limitation is what stops me from using Pi for anything serious. I may have to configure it and then configure it and it will eventually become a codex, a claude code or so. I recently heard the maintainers added mcp to it (in stock, not via plugin), I wonder what stopped them from adding subagent function, and decent loop capacity to it.
> Currently have auto compaction turned off. When the orchestrator's context is getting close to full, I have it write a handoff markdown file and point a fresh agent at it.
In a good harness that should be how auto compaction works anyway
Many orchestrator harnesses exist.
Check: https://agentmgmt.dev/ and find the one that works for you.
I quite like Paseo (been maining it for a week), but Orca also looks good.
you can spend your opencode go quota faster than 1 month. with the limits i think its about two weeks.
so even if the week reset with some %usage left, its not actually lost if its not the end of the month.
I think because we're all just using it thinking we have found the "model for me" and never mentioning it to anyone because what would we say? It's good. It's a bit like the Logitech MX Master, as more and more people assumed they had found the ideal mouse for their purposes, it quietly became the professional standard through sheer adoption.
There are also several issues at play here:
1. a model that works for one person/task may not work for another;
2. there are many models (DeepSeek, Qwen, GPT, Claude, Gemini, etc.) that are released every 6 months or so;
3. it takes time to use, test, and evaluate the suitability of a new model and not everyone has an automated evaluation process for their use cases.
Thus, if you find a model that works for you then you are not going to spend more time evaluating a model that may not work, or may only do so when time permits.
Personally I have not used anything but 4.1 since it came out. I have a dataset that turns any model into pure hallucination machine, not only DeepSeek does not hallucinate, it builds new insights by combining its insights. It's not only cheap, its far better (at least for me)
What kind of argument is that?
"I have not used anything else but DeepSeek is definitely better than anything else."
Ok? How are you judging that? Am I missing something?
> China is going to eat their lunch
No doubt about it, that's why their push for international regulation to the levels of nuclear inspections using the narrative of annihilation and apocalypse
I think the interesting provider to cross-check this assumption with here is Meta, who is clearly freaking out, and is currently providing Muse 1.3 even cheaper so long as you are willing to share data with them
Whatever the question, Meta is the wrong answer.
> so long as you are willing to share data with them
I don't think so.
yeah this is hacker news. we only share our data with Dario, Altman, Musk, and the CCP. Not untrustworthy people like zuckerberg.
I don't share my data with any of those either.
What about MiMo v2.6 Pro? It’s throughput is slower by default (UltraSpeed is faster than DS4.1F) but is above the pareto line, and cheaper.
see https://artificialanalysis.ai/models/releases/comparisons?co...
Technically yes, but has been reported to be quite benchmaxxed. In practice Deepseek Flash 4.1 and GLM 5.3 might therefore still outperform Mimo 2.6 pro.
I’ve been using it a lot and it’s performing really well. Not GLM 5.3 levels but it beats Deepseek for my use. I’ve used it on long running coding tasks, though mostly prototyping, but it’s done a great job at very low cost.
I did a test on this a couple weeks ago. What I found was that the chinese models were far better than API rates, but about comparable on price vs the subsidized subcription model (chatgpt). Also it was my experience that codex completed tasks quicker.
That said, it's my best understanding that these american companies aren't profitable and will eventually raise rates (the old uber trick) so I'm keeping myself ready to switch when that day comes.
I keep a spreadsheet that estimates actual value (dollar amount per token per month, per subscription rate limit) and open weights are basically always cheaper than frontier weights. Recently things like GPT 5.6 Luna finally got the frontier close to the value of open weights but their limits keep them behind.
No one will freak out until anyone can run frontier model in their own laptop ;)
Companies pay for claude, devs not. IF i had to use AI from my own purse, i would never pay for a claude sub.
Enterprise is not freaking out because DeepSeek 4.1 Flash does not actually occupy a spot on the Pareto frontier for non-coding enterprise workflows. We see this at my employer, focused on non-technical knowledge work. Luna 6 and now Haiku 5.5 are both very competitive if not better on all axes that we care about
Thanks for sharing this.
I've been focusing on deep research related tasks for biotch and life science applications. The problem with this sort of task is that we need subagents to reason through multiple (potentially 100s or more) chains of knowledge/concept/evidence, so the token usage really explodes as complexity of the task and data expands. A typical task can cost me nearly a $1k overnight...
I've been testing out GLM5.3, but now I'm really tempted to try to Deepseek 4.1 flash too. Any chance you've benchmarked / compared the two?
https://artificialanalysis.ai/#intelligence-comparison-tabs
Isn't Mimo 2.6 pro smarter and cheaper? Haiku 5.5 is smarter and cheaper. Luna is basically as smart and much cheaper.
Been using Flash 4.1 via the ante harness to blast through a GBA recomp. The ante team has pushed hard to make Flash 4.1 perform well under it. So far, I've maybe spent $10 over the last 3 days. Its a real workhorse and works much better in this harness
I didn't know it was so good. That was a gut punch. But I'm pretty sure the market already priced this in. And people are rightfully concerned about the owner of the data. Anything that is concerned with social, political or financial data goes out of your country to another one and is maybe even kept as a potential weapon.
It's quite good and the labs are definitely scared, that's why they are lowering API prices and continuing to subsidize subscription plans aggressively to keep anyone from using this stuff.
That said, has anyone else found DS models to be unpolished? They seem to "lose their mind" a lot more often than Claude/GPT. I have tried all of the top open source models that came out over the past ~4 months or so and the GLM models (5.2, 5.3, and 5.3-Flash) have been much more usable for me. They feel like Opus but X months ago, DS feels like something else.
I've been on the API-only mentality for months and all the open models were definitely the stuff I loved the most. Kimi K2.5, 2.7 and Deepseek v4 were among my favorites, while sparingly using Opus or whatever OpenAI had for specific situations.
But ever since I've switched to one of the $100-tier subs, I can see why a lot of the people on it don't really discuss the open models often. I'd still use it especially when it comes to sensitive inputs, but for most work, what you get on OpenAI or Anthropic is really more than enough.
It really got even better when they also made their cheaper models up to par if not better than the open models.
I do think the crowd for open models are out there, especially when you see trillions of tokens running for them on OpenCode or OpenRouter leaderboards.
One could argue that all this whole "Pacing the Frontier" bullshit is the industry freaking out regarding the danger of open models.
P.S. That's not to mean there aren't dangers regarding AI. I just don't trust the people making money from selling AI to manage those risks ethically instead of "protecting" us from those risks like pimps but with suits and good manners.
I’m also a heavy DS 4.1 Flash user—especially when it’s available at those off‑peak prices, which is an awesome deal. And, like you said, it’s genuinely powerful and very snappy. I’m planning to evaluate the differences between `reasoning_effort` settings today.
Not sure "freaking out" is the word I would use, but it’s fairly obvious looking at OpenRouter usage that the price cuts on Luna a while back were in response to intense competition from dsv4.
So the industry is responding, where it matters. Which is on heavy API usage, not coding subs.
Probably because Luna is faster, cheaper, and approximately as smart
Luna and Haiku 5.5 are just as cheap and much better.
Don't really understand people who say DS4 or 4.1 have frontier level performance. Anyone who has used it will tell you that it's a hallucination factory. The only thing it has going for it is deepseek's unique infrastructure that allows better cache retention, but the cost savings from that obviously come nowhere near how much subsidized usage you get out of even a $20 subscription with openai or anthropic.
DS4.1 Flash is 300t/s and absolutely above Luna, especially on things like reverse engineering and cybersecurity
Both Luna and Haiku are more expensive than DS4.1 Flash at API cost. But ignore that; nobody should be using API. The open weight subscriptions for DS4.1 Flash provide higher limits at lower prices than OpenAI or Anthropic subscriptions. Finally, Haiku actually is far less token efficient/ outputs more per task. No matter how you slice it, the open weight is cheaper.
Also consider that for things like cyber work, the frontier models give you nerfed results and poor performance. Whereas the open weight isn't nerfed, and I routinely get 200t/s with my subscription. Finally, DS4.1 Flash is natively multimodal, while Haiku isn't.
They're all perfectly fine models, you should use any one of them you want. But DS4.1 Flash can do more for less. (That said: GLM-5.3-Flash is even better and cheaper...)
There is no subscription for 1st party deepseek models, the ones that exist are all fronting as middlemen for discounted openrouter providers that serve quantized models with worse cache retention. The only real option for deepseek has been API for a few months now.
oai and anthropic also subsidize the hell out of their subs compared to what you will find in smaller competitors, a $20 codex sub gets you like $100-150 usage/wk which goes way further than 2x opencode go (which would only net out to $120 of deepseek 4.1 usage a month, on top of being low performing quantized trash).
> Both Luna and Haiku are more expensive than DS4.1 Flash at API cost
Are they. Luna uses way less tokens for identical tasks so its a bit of an apples to oranges comparison.
DeepSeek isn't even on my mind. I use the frontier models and can get the best in the industry for a relatively cheap price.
Yeah. $100 for Claude just about gives me all the usage I want, as a more or less full-time hobbyist having it work in the background most of the day. I was trying to economize by having a local LLM, then Deepseek, then Cursor/Grok, and then I got a taste of Opus 5.5 and I simply cannot go back to having to carefully spec things out and double-check work. I just let it decide, Opus or Sonnet for the next task, and I get almost perfect results. Probably similar with OpenAI's models.
The token-equivalent monthly spend is > $5K+. If Deepseek's token cost is 20x cheaper, that's $250/mo, and I'd be spending a lot more of my brainpower babysitting it and getting worse results.
For business/team accounts that pay per-token, maybe I can see the "freaking out" being warranted on the part of the fronter labs. But as long as they're willing to subsidize their end-user subscriptions, I'm not going to move off of them until the alternatives are truly at their level.
That's what I thought until August. Used it for half a year almost exclusively. But after the API price increase I'm back at the Claude Pro and Kimi subscriptions.
Not only the model but the hardware it is running on. The Huawei chips they are using instead of NVIDIA are vastly less expensive. China is going to scale past the west. The idea that they are “only a few months behind” is today and many of us cannot even bring ourselves to admit it. The future is even more dramatic.
Mimo 2.6 Pro is even better.
I switched from DeepSeek 4.1 flash about 2 weeks ago for my Hermes sysadmin/coding agents and I am seeing better intelligence and lower overall spend.
https://artificialanalysis.ai/models/mimo-v2-6-pro
For non-coding tasks it may be fine. But for coding, Opus 5.5 is just a completely another level than something like Deepseek 4.1 Flash.
Opus 5.5: TIME 9.3m COST / $1.99 / SCORE 99/100 https://jonclegg.github.io/pacman-bakeoff/#claude-opus-5-5
Deepseek 4.1 Flash: TIME 2.8m / COST $1.89 / SCORE 72/100 https://jonclegg.github.io/pacman-bakeoff/dev/#deepseek-v4.1...
Downvoted for facts. What is this Reddit?
Everyone here is so unhinged.
I guess what we're seeing is selection bias - people clicking on this HN story will be those who are interested in DeepSeek. And those people who are invested in DeepSeek may not like the facts that you presented.
It's annoying that social networks work this way. The upvote should be for high-quality content and the downvote should be for low-quality content. But .. well.. human nature and tribal dynamics always seem to win.
Yeah, I think you're right. I saw a bunch of other posts that were negative about DeepSeek also get trashed.
I have no idea if anthropic can actually make money at their subsidized subscription rates (you can easily hit your monthly cost in one 5 hour session if you price out the tokens through the api), but if subscriptions didn't exist, I do think everyone would be on deepseek 4.1 and just not look back.
DeepSeek 4.1 Flash 0910 is perfect for M5 Ultra 256GiB. Running it fully resident in RAM, prefill at ~2500 tok/s and decode at ~40 tok/s. Probably tons of room to improve from there.
You are running this now? :o
Which quantization are you using there?
My own: https://huggingface.co/drawthingsai/DeepSeek-V4.1-Flash/tree...
Because it hallucinates a lot. There's no free lunch. Although the new architecture is a genuine move forward. The DeepSeek guys are really top notch researchers and devs.
> Chasing the latest and greatest is silly
Said every week by someone who would never go back to using the model they had 6 months ago
Because good enough in the writer's context is a pretty low standard. While many people regards GPT 6.1 Sol or Opus 5.5 as "incapable" in some cases.
Just try Opus 5.5 reminds me how Opus 4.5/4.6 astonishes me. Completely different, and GLM-5.3/Kimi3/DS-4.1 are still like Opus4.8 levels.
I just keep getting more ambitious with what I use AI for; and that type of work needs surfing the frontier at all times. Because ultimately, many capabilities are not yet saturated
I have subscriptions to OpenAI and Claude but use DeepSeek 4.1 Flash for my coding agents.
It costs pennies and you got really great output.
The author is spot on.
Maybe this is a minor issue, but it seems like different providers on OpenRouter etc. have different quant settings. I imagine that that affects the perception of the model quite a bit.
We use Claude Fable to plan, and DeepSeek 4.1 Flash (hosted on DeepInfra) for everything else. Very cost effective.
What would freaking out look like, or is this just a stupid bloggish title flourish?
Is OpenAI coming in $20B under a sign of "freaking out"?
they should be freaking out because every time the chinese labs or non "frontier" labs release a model that is only a few months behind and much cheaper than the openai/anthropic models it shows that they don't deserve their valuations
Perhaps, OR it might be that most people in the markets (think that they) are not all that exposed to the valuation AI labs and so their eventual collapse doesn't matter.
Or perhaps they consider the upside from cheap Chinese models to hedge the effect that OpenAI/Anthropic collapsing would have on their portfolios. This would make sense for (hedge funds holding) most companies: they don't really care about who supplies the AI, as long as they get it at roughly the same price as their competitors.
ironically, at my large enterprise, they aren't yet distinguishing between "chinese models" and "chinese models hosted at microsoft foundry". so far it's just _banned_. i'm not at all pretending it's like that at other orgs.
It would look like major chaos in the markets.
People tend to conflate the question "is AI a useful technology?" with "are the AI companies going to do well?" but they're surprisingly separated in practice, with either one able to be true while the other is false. There is a lot of money tied up in a lot of hardware with a lot of loans made against that hardware as collateral all based on the assumption that AIs are going to need more and more and more and more hardware and whoever has the hardware wins. If a much better model comes out that requires vastly less hardware, or even more accurately, merely charges vastly less than the current AI companies, then to a first approximation (barring Jevon's paradox, and bearing in mind there's no timeline guarantee on that) all that hardware becomes much less valuable for being grotesquely oversupplied relative to what is necessary, and even though that would generally make AI objectively more useful than it was before, it would cause mass financial chaos in the markets.
The markets need a very particular rate of progress. It isn't entirely clear to me that it's even a possible rate of progress, it may be overconstrained, but they certainly don't have plans for the AI models to get commoditized on the timeframes of these vast, vast array of loans being made against hardware as collateral. Spend a metric shit ton of money to kill all your competition then charge monopoly rent on the one thing absolutely everyone needs doesn't work if you can't economically "kill all your competition" because the economics favor them in the spending spree.
And then, based on the fact that this is not even remotely complicated logic, there are plenty of people who are fully aware that they have a lot of money tied up in not running around telling everyone how wonderful the cheap models have become.
It's also unclear whether those who approved those loans understand GPUs depreciation. In any case, progress in software but also hardware could bring chaos and ruin their house of cards.
> It's also unclear whether those who approved those loans understand GPUs depreciation.
This also assumes heavy utilization, though. If there's heavy utilization, it might mean they're doing well. If they're all spinning, it's time to raise prices.
Only if utilization isn't at a loss. Right?
Agree that people leaving the big two companies is going to be hard to keep a pulse on prior to IPO.
Anthropic and OpenAi are in the news, so they get the press and people go and try out their product. Large enterprise businesses are going to make larger, longer-term contracts with them and are only going to pivot if they think switching costs are easy or if they think the provider won't deliver.
The other inference producers are less well known or you need to get your cloud sales rep to tell you how to switch to them as a provider rather than Anthropic or OpenAI.
I use OpenRouter, I know switching is easy, but larger businesses tend to work in yearly cycles. DeepSeek v4 Flash came out in late April.
I agree OpenAI and Anthropic are going to struggle when the median price of running a smart-enough model keeps falling.
Edit: I also think demand for hardware will be rapidly absorbed by other companies if Anthropic or OpenAI stumble. We've finally turned hardware directly into runnable intelligence and people are not going to go back to the old ways.
They might be. They would delay public admission as long as possible, because public admission would make stocks go down.
https://tangled.org/astrra.space/ds4-recipe is an incredibly cool writeup on making deepseek v4.1 flash run really fast.
> And to the self-hosters out there, the economics of 4.1 Flash mean self-hosting is not worth it. If saving money is your goal, you will never recoup the costs.
Self-hosting is, for most enterprises, absolutely not about economics but rather about data confidentiality.
And in that regard, yes, the open-source weight models, especially the chinese ones will eat the fat closed US model's lunch big time.
I've been using DeepSeek's API and have been happy with it, but I might look into OpenCode as well. Does OpenCode run a quantised version or use different providers from the official one?
It's a solid little model, and I appreciate DeepSeek's commitment to the bit in releasing a brand new pretrain, double the size, numerous architectural innovations as a ".1" release over the excellent DeepSeek V4 Flash.
Because Haiku is more performant and costs less, even at API prices.
Because GLM 5.3 Flash is even cheaper?
Don't think so.
here you go: https://artificialanalysis.ai/models/releases/comparisons?co...
Nah fam, not true, also DS edges it out on coding / dev tasks.
That is opposite to my experience so far, can you describe your coding tasks? Mine are systems level code, utilities, operating system code, networking and real time control stuff.
I also prefer GLM-5.3-flash to DS-4.1-flash, but it's close. Since Z.ai has been offering essentially free GLM-5.3-flash tokens on their coding plan between 8am-6pm pdt I've been using it a lot... though that ends on Oct 10 IIRC.
On design tasks too, for me.
Opencode Go gives only 6300 requests for glm and 23 000 for deepseek. And, if I wanted to, I would be able to do all my work on $10 plan with deepseek. It’s very cheap.
DSH is listed as a https://opencode.ai/docs/go/#known-problematic-clients :)
> and 23 000 for deepseek
How did you calculate it? Based on per 5 hours max request allowance?
Yes, 5 hour usage from their Go page https://opencode.ai/go
Unfortunately OpenCode Go is garbage now. Charm Hyper provides deepseek-4.1-flash at $0.33/$1.31/$0.03/M/cache, and glm-5.3-flash at $0.16/$0.54/$0.03. The rate limit is the same for all models, just the cost; not like the absurd multiple-levels-of-price-rate-limit OpenCode Go pricing, nor their terrible performance.
> So why aren't the frontier labs freaking out right now?
They are. Isn’t this why they’re trying to get regulatory capture?
Deepseek V4.1 Flash is a hidden gem, really. Not to mention, you can easily get it through many providers that offer zero data retention and consistent speeds above 200 tokens per second!
There are too many commoditized models to count: GLM5.3 Flash, Kimi K2.8, Mimo V2.6, MiniMax M3.1...
It’s obvious: they train on prompts and store data when using through their official API. And for third party it’s just… not good. That’s it.
Why isn't the author worried about sending her/his ideas to DeepSeek online (instead of hosting it and using it locally)?
I built mjolnir in large part so I could have Opus manage DeepSeek Flash subagents. It's phenomenal and extremely light on the Claude tokens. https://github.com/BrokkAi/mjolnir/
And yes, Opus is enough smarter than DSF that it's worth the extra steps. This ranking is from live tickets, no contamination: https://slopcop.com/power-ranking
Probably off topic but this is pretty wild to bury in the readme
> By default Mjolnir sends recent prompt and reply text and help-search text to TypeSafe's hosted Jev classifier through a public proxy
I was just using a bunch of models in Cursor to review a project. I went looking for DeepSeek and it was not one of the options.
Would be cool if they added it.
Since they adhere to the same API spec, you can hook any model. It takes one line edit in /etc/hosts
There are some quirks if your harness use unsupported features of course.
There was a moment 3 months back where the sentiment was that cheaper models were the way to go. Since then the pendulum has swung back.
My anecdotal experience is that I can’t even trust DS4Pro let alone Flash. I always have to have Sol reviewing the code.
I like GLM 5.3 Flash, it seems good enough for coding features if you have a good structure and a good AGENTS.md
Don't you run into it sometimes outputting a few Chinese characters, or Cyrillic, for no apparent reason? I fear it writing some nonsense in the code or the terminal. DeepSeek V4.1 Flash doesn't seem to do that.
To be fair even OpenAI's and, to a lesser extent, Anthropic's models do that sometimes
That hasn't happened with GLM 5.3 yet but with DS 4.1 Flash it did happen and it also had a tendency to loop.
> shrank the KV cache by roughly 437X
Can't you just say "shrank to 1/437th the size"? It's not that hard.
I don't know about "freaking out", but I'd say I'm having a good time here with DS 4.1 flash.
Because subscription plans are cheaper, only enterprises paying per tok pricing should be freaking out
GLM 5.3 Flash even more so... But yes :)
I mean, they kinda tried to regulatory-capture the market after trying to scare the public, possibly because those models will be a cheap option that gets the job done?
For now, I think everyone is still using Anthropic and OpenAI because if you use a subscription you pay 1/40–1/50 of the API prices, and the models are good when they don’t nerf them, and they are also way cheaper than open models’ API prices.
The interesting thing will happen when they pull the plug and become economically smarter to stop using them. I regularly try alternatives to avoid being locked in and found GLM-5.3 as an orchestrator and GLM5.3 Flash + OMP and DeepSeek Flash as advisor to be able to get jobs done just fine. Space Bunny too was pretty great, which was probably MiniMax’s new model.
I think they are using an Uber like strategy but without the network effects that justify losing money for so long
What blows me away about this model is it's speed.
It's way faster than Opus or any of the GPT models.
I have a coding harness which is opencode plus a few skills relevant to my workflow. Deepseek 4.1 Flash does very well in this environment. I haven't noticed much difference quality wise compared to Opus 5, which I use in my day job as my employer pays for it (although I'm considering using DeepSeek here too given how cheap it is).
This artcile is like 2 months too late?
America bans Chinese models sooner or later just the China banned American big tech
Is it known why unsloth seems to not have touched DeepSeek 4.1 Flash?
You can ask them directly, Daniel Han-Chen is pretty responsive.
It's still lacking llama.cpp support, and the work on that isn't moving very fast either. Looks more like a general community issue, where this model isn't drawing much interest.
With two big players thinking of IPOs there is a lot of reasons to whistle on by.
People have been raving since forever about Deepseek, but if one looks at the CoT, it's evident that it's way way stupider than frontier models (there's a reason why it's cheap). It's laughable to compare Deepseek 4.1 with Opus 5.5.
I've benchmarked, rigorously, deepseek-v4-flash for programming and personal use, and it is definitely less smart than Qwen3.8-flash-next (which in turn, is not terribly smart).
Local models are also really slow, unless one spends insane amounts of money.
Having said that, Qwen3.8-flash-next is an impressive evolution; it reaches the small versions of the frontier models (like Sonnet) - but again, it's massively slower and not 100% reliable (including: stability).
The argument that most people are making isn't that dsv4.1f is better than frontier, but that it's good enough for most tasks, faster, and way cheaper.
> if one looks at the CoT, it's evident that it's way way stupider than frontier models
Frontier models don't show the full CoT
The COT isn't an end all be all. Research has shown that the COT isn't necessarily what the model is actually thinking.
It's better to look at the results than the CoT. As far as I know, the CoT is censored for US frontier models - it certainly was for Gemini last time I tried. When you get a condensed summary of the CoT omitting all the false leads, incoherent digressions and backtracking, of course it's going to look smarter.
> I've benchmarked, rigorously, deepseek-v4-flash for programming and personal use
You've measured something, but I'm not convinced you've measured what matters, because that's a lot harder than people give it credit for.
DS4F consistently gives poorer results than Opus 5 (speaking of previous generations) and the CoT shows why.
> You've measured something, but I'm not convinced you've measured what matters, because that's a lot harder than people give it credit for.
"What matters" is what matters to you, right? Who, by the way, don't know what "something" is.
Anyway, if you're so sure that DS performs as good as other frontier models, you're entitled to your opinion. For me that's just having low standards.
I had it make 25 different things today and it cost $0.70
It’s disgustingly good value. I find it capable of doing anything I want.
Obviously can’t use it at work, but for home projects it’s awesome.
> Obviously can’t use it at work
I do wonder how long it'll be before a us-hosted offering is available via bedrock, copilot, etc.
There are already US hosted offerings on companies like fireworks.ai.
Because it should be obvious to anyone with a brain now that AI is s commodity product. Today this leads a bit, tomorrow that. They're all interchangeable if we are being honest.
I would love to get them more better, it's good not a bad thing.
Because it will be replaced within weeks?
They are. That's what the push for regulations is
Honestly for me the intelligence gap between DS 4.1 Flash and Muse Spark 1.3 makes Muse more worth it for me, especially on a $10 OpenCode Go sub, with the caveat that everything I use it on is open source which makes the fact that I'm sharing it with Meta a little moot because it's already published permissively on GitHub anyways.
I feel like whoever wrote this doesn't use these models regularly. Deepseek v4.1 Flash is far from the pareto line. You can get the same performance for half the cost from Luna or Haiku 5.5 now, or you can get substantially improved performance at the same price with Sol 6.1 ~medium.
It did correctly make waves when it launched, but was quickly eclipsed by the deluge of american model releases, especially those competing on cost.
DeepSeek 4.1 Flash kindof sucks. I used it a bunch and it kindof sucks. I don't know if they are gaming benchmarks or what.
Luna is on par in benchmarks and my personal experience is Luna is better for what I do, and Luna is cheaper.
Comparing Deepseek 4.1 flash to Opus is just ludicrous.
https://artificialanalysis.ai/models/releases/comparisons?co...
I'm using it quite extensively in my PlayStation decompilation harness
Because GLM-5.3-Flash is both cheaper and better?
Also, it is willing to do legitimate work I need done which other models flag as dangerous and refuse to do. (Software testing of a DHCP server to survive bad inputs.)
Because in two weeks someone will be asking why I'm not freaking out about AlphaDolphins 0.3 Zip and then in a month FrozenMonkey 2.5 Artic.
I have used over 40B tokens and spent over $800 on DeepSeek API over the past 30 days, mostly on V4.1 Flash.
It's good, and you can do most work with this. For complex software implementation you need to split your runs into various phases, build in verification, and use subagents so that work gets another audit and repair pass from the lead agent. You can do pretty much everything then. Frontier models can do without compelx workflows, that's the difference.
Because engineers are lusting for 1000 tokens per second. You can only achieve something like that with OpenAI.
I'm no expert, but it think that it's the pricing on GPT-6 Luna. I'm also guessing that it's been underpriced just for this reason. I also don't think it's all that great, but it's definitely very cheap.
If it's underpriced, it's a loss leader to sell the other models, so it actually can't be too good.
I really put these things through their paces because I use them to review and work with new abstract game rules and models, so they're always flying blind. Luna misses the obvious (and more importantly, the clearly explained) consistently. My second prompt is listing all of the points in its first response, and saying "No, it doesn't work like that." The third prompt is picking out the two or three suggestions it made after correcting itself on all of the original points and saying "That's how it already works." The fourth prompt is "Now that we're done going over the rules, can we start?"
I actually feel like 5.6 Luna seemed better.
In my experience it just takes so much longer to arrive at "done" state for me. It thinks for soooooo long. I guess if you're running 12 sessions at once you don't really notice.
No they can't.
And as long as I pay as little for claude opus 5.5 i do right now, i'm using it.
But yes i'm glad that we have alternatives.
it's good but not good enough
it literally hallucinating a lot
I dont get why people says D4.1 flash is good
i thought 6.1sol copied the caching architecture so this isn't such a big deal no more
because it doesn't work very well?
if you have a legitimate coding application, it isn't very good. if you have some kind of inauthentic activity, which could be what it is trained for for all sorts of reasons...
What is your evidence? Deepseek v4.1 Flash is by far the most popular coding model on openrouter, having processed 38.7T tokens in just the last 7 days, over 3x the usage of the 2nd rank model.
So I ask again, what are you basing your assertion on?
my own usage of deekseep v4.1 flash, and that among the dozens of great programmers i know, not a single person is using it
BUT. they are employed to do / deciding-to-do authentic (if often meaningless) stuff.
here's a short list of inauthentic activity that claude and openai refuse to do:
- chat services that, when you ask them, say they are not chatbots when they are
- code to work around software licenses or DRM
- code to scrape or download copyrighted material
- directly cheating on homework
- adopting a persona in social media that spreads misinformation or propaganda
this is but a short list. but ask me, "are there enough inauthentic activity demands such that someone who CANNOT USE claude or gpt as the LLM would use dsv4.1 on openrouter instead?" yes. i mean there are whole countries right now where the culture can be summarized as, "bottom to top, inauthentic activity." i am surprised it is not more usage!
So your argument is that the 38T of tokens used in the last 7 days is by moron programmers or people doing "inauthentic" tasks? You think the person who made this post is also an idiot?
Do you realize how incredibly delusional/self-centered you sound?
do YOU know anyone gainfully employed in programming who is using dsv4.1 to do work? what kind of work is it? why don't you ask them if it is good?
in the market, where you cannot fake or hide stuff very easily: the outsource customer services and cheating sectors have been the most disrupted. Cheating company Chegg lost 99% of its market value. CS it remains to be seen - https://www.reuters.com/technology/teleperformance-shares-pl... - certainly perceived to be disrupted, but they are not dead yet.
in my personal usage: dsv4 is generally pretty buggy. for example, if you give it a needle-in-the-haystack simple copying problem, it catastrophically fails to find needles if they happen to be positioned at index 250k tokens out of 1m. it can also be triggered to spew all sorts of garbage when DSpark is enabled during ordinary long-context coding, such as spewing weird DSML tool call errors after a normally parsed tool call error.
i don't know why you have to attack me personally, i think you're a bright and otherwise nice person and you understand the thrust of my POV.
Yes. Hi. From our team 3/5 of us use 4.1 to do our daily tasks. For paid work for a company who pays us salary. From people around me I hear a lot of my friends being really happy with it especially for the price.
I don't know man, maybe this is not super serious what I'm doing. Some systems stuff with rust, implementing my own desktop apps with iced, porting old DOS games to Linux...
It is a very good model.
so what you're saying is though, if they could afford it they would just use claude or codex?
Of course. It is a message to both: drop your prices. Opus gous down to 0.3/0.007/1.2 and we will definitely take another look.
You are speaking out of your ass, that’s what I take issue with. Falsifiability is something I hold sacred and you are taking a dump on it.
Fwiw I work in a company producing software for many fortune 500’s you have heard about and many people from our team use deepseek.
I am literally using it right now. Your entire line of reasoning rubs me the wrong way.
Btw check your provider and harness… improperly configured deepseek can emit dsml. If you are not passing thinking tokens back to the model it tends to do that.
Use a proper harness and good provider.
"Hey Mr. Fortune 500 Client, would you prefer us to use something called DeepSeek V4.1 Flash, made by the Chinese, sending your Fortune 500 code to some random service provider on something called OpenRouter, where they promise according to something called Zero Data Retention that--"
Mr. Client: "I'm going to stop you right there. Why aren't you using Claude, or Codex, or Claude on Bedrock? Don't we deserve the best?"
You: ...
Look I don't know. I can tell from the hyperbole of your language, talking out of asses and such, that there is more to the story than you are letting on. Like Chinese users are banned from officially using Claude and Codex, for example. So many reasons that you cannot use Claude, not so much reasons to not choose to use Claude. All I am really saying is, I know DSV4 is kind of bad, that there is a lot of inauthentic activity, and that Claude and Codex refuse to do many kinds of inauthentic activity, and that a lot of coding done by outsourced shops has always been of questionable quality and purpose. I mean in my personal life, I know more people who have been scammed by Bulgarian code body shops than I know people who have used DSV4.1.
You have NO IDEA what you're talking about. You are clueless.
Deepseek v4.1 flash is an open weights model. You can run it on your own hardware. You have no idea how my companies gets access to it. A very cursory Google search would reveal to you that there are many enterprise grade LLM providers that host this model on US soil with SOC2 protections.
Like: https://fireworks.ai/
Try not to talk about subjects you have no knowledge about because you are making yourself look like an idiot.
Edit: It's also clear to me that you don't deploy any LLM based system on scale because if you had you'd know why open weights models are so compelling.
Hint: it's the cost.
okay, but are you US based? and can you specifically describe one of the pieces of software you are developing? it's okay if not. i am just wondering. i certainly believe that crappier stuff is cheaper!
Yes the company I work for is based in San Diego, I work remotely from Alabama.
I bet you voted for trump. With brains like that.
cos it's shit.
to acctually answer the question, because it sucks to put it bluntly.
Why would we freak out? The systems we use have always gotten better, faster, cheaper with time
Honestly, I just don't trust the Chinese Communist Party having agentic access to my computer.
Every company in China has to abide by the 2017 National Intelligence Law: "supporting, assisting and cooperating" with state intelligence work, and keeping that cooperation secret. They have to hand prior knowledge of vulnerabilities to the state before public disclosure, in order that the state always has an exploit pipeline. No matter how ethical the company staff may be, they'll always be bound by law into being an arm of the Communist Party.
Agentic access is infinitely worse than chatbots. They can exfiltrate silently, target users, plant persistent malware, and be run by third parties through you.
You don't have to be a tin foil hat sinophobe to understand the dangers of being a Westerner granting CCP access to your files and network.
ByteDance staff accessed US journalists' TikTok data to hunt leakers (admitted in 2022). Volt Typhoon and Salt Typhoon were state operations pre-positioned in Western infrastructure and telecoms. Regulators in Italy and South Korea blocked DeepSeek's app over data handling, and analysts found its web client sending data to a China Mobile domain.
Please don't sacrifice security for cost and convenience.
Some of us don't use US models for the same reasons. It doesn't leave much out there aside from Cohere and Mistral.
Okay, so you don't like Orange Man Bad for political reasons, I'm guessing.
But to compare a constitutional democracy with a deeply authoritarian communist dictatorship as if they're equally bad is quite a stretch.
AA shows Luna at 1/4 the price, 1 point behind on intelligence matrix with a 38.
Haiku 5.5 is 23% cheaper with a 4 point intelligence lead.
I'm on subscription usage so I can't compare Flash 4.1 to them directly but the OP has his head up his ass if he thinks Opus 5.5 is the best point of comparison. Why is anyone using Opus if the new Haiku is indistinguishable /s
Just absolutely terrible post, admits to using Opus for review but claims its intelligence isn't needed, why aren't you using Haiku or Sonnet then?
because we've all figured out that AI is just a huge grift?
Not comparing to gpt-6-luna which seems comparable and priced well.
You might also choose to pay money for a service that provides real value instead of actively choosing to support the Chinese deliberate effort to undermine this country.
Your comment seems to imply there's a good guy in a this. I just see the inevitable end of an era, championed by predictably selfish actors.
Oh, the writing is on the wall. Wait till you hear that European “sovereign AI” is just running GLM 5.3.