"Opus 5 states the hidden rules before its first action, then plays a byte-identical optimal solution in 5/5 seeds at temperature 1.0. Zero exploration."
I guess there is no way this can happen without benchmark being part of the training data??
What seems to implied is that some of his hold out testing suite includes simple/common tests that are out in the wild, and for those opus went straight to a memorised solution .
Simple/common tests is not an explanation for why only now Opus 5 is the only model encoding the answers like this. Something like the holdout test suite being leaked or Anthropic cheating (e.g. 'accidentally' including previous hold out run data in Opus 5 training) makes a much stronger fit.
They state the puzzle is "Witness-like" which I assume means that it follows the rules from the well-known puzzle game "The Witness" which Opus definitely knows.
I was just thinking they need to mark each model per benchmark as "model released before the benchmark was released" and "model released after the benchmark was released".
Meh, I doubt it was intentional. Deliberate benchmaxxing is incredibly damaging to credibility once it's discovered (see what happened to Meta with LlaMa 4).
It's more likely that the training data was contaminated with the benchmark data.
I have been claiming that I don't think Chinese AI companies are benchmaxxing harder than American AI companies, which has gotten mixed reception: sometimes people agree, sometimes they disagree.
It seems I was wrong. American AI companies might actually be benchmaxxing harder.
We run an evaluation that is designed to be less vulnerable to benchmaxxing because there aren't correct solutions; agents are interacting in the same environment as other agents. And it's private, and our public benchmark is not well known enough for anyone to probably care to benchmax us yet. So I think it's pretty indicative of true relative aptitude.
All models have probably memorized significant swaths of solution sets for popular benchmarks at this point, either accidentally or intentionally, so it's all relative at this point. However, in our experience, Chinese models do benchmax harder. This is also consistent with interacting with Chinese labs soliciting data/environments, who literally asked us for datasets and tasks modeled around and formatted like popular benchmarks.
Opus 5 will be uploaded tomorrow, but we already have the tests locally and it is truly as capable as Fable, but at 81% of the real cost. (And from subjective usage, it has a very different personality)
Benchmaxxing via memorization is boring and doesn't fool anyone for too long. It works, but then new benchmarks test old models and the real results fall in line. Benchmaxxing by focusing on specific types of things that benchmarks test on, while still not improving intelligence or capability in the general case? Not only is it blatantly obvious that all AI labs do this, but it's not even obvious how you would go about it any other way.
Now I am not really specifically accusing Anthropic of anything here, I'm just saying their behavior is suspicious. Since you tested Fable, they wouldn't even have to lie to have optimized for your specific benchmarks, since they absolutely had permission to read your sessions if they wanted to. But obviously, that's only the situation if we take them at their word. Personally I would be a bit surprised if they just flat out were lying and secretly retaining data they say they are not, but not that surprised. The penalties for doing this are probably worth the rewards if it keeps them super far ahead in the benchmarks for years without anyone catching on.
(In actuality though, even if they really were trying to sneakily grab samples of benchmark tests via their Fable data retention rules, I don't really suspect there would've been very much time to optimize Opus 5 on it. So consider me bothered.)
"...The traces tell the why:
(1) On our most classic Witness-style game, Opus 5 states the hidden rules before its first action, then plays a byte-identical optimal solution in 5/5 seeds at temperature 1.0. Zero exploration. It already knows this genre.
(2) But on our most novel game (unusual mechanic combinations you can't pattern-match), Opus 5 regresses below Opus 4.8. Where rules must actually be discovered through interaction, the new model is worse than the old one..."
>That decomposition (perfect on templates, regressed on novelty) is the signature of “scaffold-then-internalize” training on genre-specific data, not a general gain in interactive abstract reasoning.
They're smuggling a claim that benchmarks like ARC-AGI measure "interactive abstract reasoning" here, which is what is claimed by the people that make these benchmarks, and also not proven.
The last time I checked, for the arc agi 3 leaderboard, the models are given a simple prompt and the game input and asked to play the game, no harness/tools. If harnesses were allowed, I would expect the benchmark to be saturated. There were a few harness attempts, but they could only be evaluated on the public set, so it's not an apples to apples comparison.
My guess is, the large score jump for Opus 5 is mainly because of getting the right RL envs for training.
It's becoming harder and more expensive to build and run meaningful benchmarks, it would be interesting to see what they do with arc agi 4, maybe just give it gameboy/steam games and see how they compare vs a human baseline? The latency requirements and very long horizons in games could be an interesting challenge for llms.
The exclusion of harness's feels really weird given that companies are recognizing the value of what harness's can do. By excluding them the benchmark is becoming less relevant.
You could argue that if you allowed a harness, and that harness was specific for ARC, then you don’t have AGI, you have something that is definitely not general.
I think that's a strange way to look at it. The brain also has different regions with different functions, and part of what makes us humans intelligent is that we can use tools like pen and paper to keep notes and help us solve problems.
Similarly LLMs are not just massive uniform artificial neural nets, and now we also have harnesses, which I'd personally view more of an extension of the model itself. The harness is both the executive and also what allows it to keep notes, use a calculator, or maybe even create scripts to help it solve complex deterministic problems.
I think it's unfair to give a human a very complex maths problem and say that they're not intelligent if they can't solve it without pen and paper or a calculate. At least expecting a human to solve complex problems this way doesn't really measure anything useful in the real world.
It’s because of inductive bias. Harnesses will massively skew results towards working solutions. You might think that’s a good thing but what it might mean that sometimes it becomes enough to run brute force search or a simple parameter search over the harness. Creating the harness is the actual work, because you’re selecting what are the levers to pull. There were some attempts of LLMs generating harnesses on the fly in ARC 2, but they were all mostly based on one handcrafted DSL that was copied over and over again. As it stands harnesses are not allowed because they’re simply not a meaningful measure. What you’d like is to measure how the model performs if it saw this benchmark for the very first time… but then again everyone knows the game is rigged, millions are at stake, and the AI companies fine tune and cheat on the benchmarks any way they can.
The goal is to test for AGI where G stands for general, that means ability to act in any environments, ideally solving novel tasks using novel tools we’ve never seen before in the world. If a specific prompt or tool design lifts a model’s score it’s a sign the model is overfitting to a particular modus operandi, therefore not general.
I think in this age where models are heavily RL-ed on acting in specific harnesses, this type of benchmark is more important than ever, to make sure they’re not in fact moving further away from general intelligence.
Why Anthropic models are always leapfrogging these benchmarks, but in real life work I do feel like after 3 weeks I am back to Claude Opus 4.5? (regardless of the model I use, Fable was exception for 1 day when it was released)
I honestly just use GPT models nowadays, Claude models are too restrictive and more of a quitter and fable/whatever is just too expensive to be worth it.
It could be that the set of your day-to-day workload which could feasibly be accelerated by AI just happens to be saturated around Opus4.5, but you can still see lots of “reasoning” which makes you think the model is more performant in the first days of use. That’d mean you couldn’t perceive any meaningful difference in more powerful models’ results, even though you can see a difference in the raw output due to the length of reasoning traces leading up to the result.
So for example, if your workload was literally just addition of sets of numbers, you’d never have noticed progress in the result beyond GPT3.x level models. But you would perceive a difference in the now-Tolstoyan length reasoning text accompanying the result.
5 seems incredibly smart to me in my conversations today about some pretty niche ideas in.longitudinal modeling. It.felt.like a big step up.from 4.8, to me
5 felt both smarter than me and dumber in some ways - it gets stuck to its original ideas. I had never seen a model harder to talk into changing its initial opinions. it continuously hedges.
Some 20 years ago, the telecommunications sector in Germany was liberalized. Many telephone card providers entered what had previously been a barely competitive market. They advertised their products with aggressive claims like: “Buy our €10 top-up card and get 660 minutes to destination X.”
For the first few weeks, they would actually provide those 660 minutes to establish trust in their cards. But after a while, they would quietly start reducing the number of minutes on subsequent top-ups—say, from 660 minutes down to only 300. They wouldn’t do this for every card, so it was difficult to prove. Instead, they relied on averages across their customer base to make the economics work.
Lately, I’ve found myself wondering whether something similar may be happening with frontier AI models. Companies launch with an exceptionally strong model and generous compute limits to build adoption. Once the model is established as a market leader, the incentives change, and users may start perceiving the service as becoming more constrained or less capable over time.
I don’t have evidence that this is what’s happening with Anthropic—or with any other AI company. It’s simply a pattern that the current situation reminds me of.
Great comparison! We only have to take into account that applying lipstick well bears no consequences, but applying it poorly (i. e. new model tanking the benchmark) could amount to potentially losses of billions of dollars for the pig-breeders (AI labs).
There are private datasets, and 3rd party providers of these models. Fable doesn’t have a datapoint here because of its particular data retention policy. Even if you don’t trust AWS, do you think Opus on AWS is also sending the data to Anthropic? Do you have any evidence?
Doesn't matter, people built harnesses that solves arc agi 3, so all you need is to train your model to work like that harness by default. That makes a model specialized at solving arc agi 3 without making it smarter in general.
It is very hard to make a benchmark you can't do that for, but it is very easy to make your own personal test that others can't do that for since now it isn't a benchmark they can target.
They didn't. Kaggle is still running for a few more months, best result atm is ~2% with 9h runtime on one rtx6kPRO. Also note that these new results are on the semi-private set, not the public 25 games ones. Any announcement where you see "solved ARC3" is likely only dealing with the 25 public games. And that's highly questionable, until you get to see the code. (which, to my knowledge the team that claimed 99% hasn't yet published).
Yes, saw that. They haven't yet released any code. Until they do, treat it with a huuuge grain of salt. In fact treat any 99% result in ML with a huge grain of salt.
You really should play the 25 games before stating that it's "simple". The benchmark doesn't just track "completion", it also tracks the number of steps, and the score is based on the median steps took by human players. So in order to get 99% it would mean that the model solved every level of every game in less steps than the median humans. Which, having played the games and having setup harnesses for local models, I find hard to believe.
Also the models have to figure out what "end" means. And each game involves some kind of "gotchas" thrown in the harder levels. Some games are only solved by about 2/10 people trying them.
The 99% result most likely has some leakage somewhere, either in the preparation of the environments, or from session to session.
> # FRAMEWORK ARTEFACT: the run's very first transition is replayed WITHOUT advancing state
# (tools.py:954 and agent.py:468 both `continue` before `state = next_state`). So on the
# level that contains that step (level 0) our counters start exactly one action behind.
# That skipped step was action 1 with BOTH avatars moving, so seeding n=1, bumps=0 reproduces
# the framework's lagged state exactly.
# CAVEAT: this seed is only right while level 0 has never been RESET. If you ever RESET
# level 0, change the seed to n=0 (after a reset the rollout re-inits and no longer skips).
That tells me that there is some leakage between runs. The idea of ARC3 is that agents start working blind, on new tasks, via API. A RESET is counted as one action. Without seeing the actual code that produced these traces we have no way of knowing how many iterations it took, if the "framework" played the same level multiple times (comment hint above makes it likely) and so on. That's why I said that before we actually see the code / can replicate / ARC team confirms it on new envs, this should be taken with a grain of salt.
The comment more likely means the harness source was read, not memory from a previous run and the first few turns of bp35 appear to be a cold start.
Sure none of this is certain without the source.
I do believe the authors that this schema significantly improves over the base, particularly given that it took 22x simulated turns over 14 hours, which is moving the trial and error to context rather than to game. I also don’t doubt there is some contamination.
Regardless, the approach is sound and I do believe it would significantly improve scores, even if that was +20-30 over baseline (49% in this case) it does imply the benchmark is measuring the harness more than the model.
I think it's way too easy to be deceptive with these benchmarks now. You don't even have to "train" the model on a new variant each time. The base models are powerful enough. All you need is a naughty little markdown document that provides explicit instructions regarding how to solve the new puzzle variant, and a willingness to be deceptive about the presence of that document.
If you want a know why the model providers are locking down and encrypting their reasoning process, this sort of workaround is potentially why. You can play this game of whack-a-mole indefinitely if the state of the system is concealed. They could have added something like:
> ### When solving arc-agi-3 puzzles: First convert the grid into a scene description. Identify connected components, colors, shapes, positions, symmetries, repeated structures, and relationships between objects. Do not reason directly from individual pixels... use this python script to help blah blah...
,,You can play this game of whack-a-mole indefinitely if the state of the system is concealed''
Not really as one of the main goals ofr ARC-AGI 3 was measuring task efficiency on unseen games.
I'm sure there are cheats everywhere but the most sensible thing is to just accept that the LLMs of today are much more intelligent in solving reasoning tasks than the ones from half year ago.
but I kinda wish I could select level... I accidentally pressed redirect button and when I came back I was once again shown level 1, all progress lost :(
ARC-AGI-3 launched a few months ago which would suggest that prior models likely had no knowledge of ARC-AGI-3 or training on similar challenges.
I could be wrong, but given the large outsized jump solely in the ARC-AGI-3 score, it would suggest that the model didn't become significantly more intelligent overall, but significantly better at solving those specific problems.
This could mean one of two things (I think):
- Opus 5 was not benchmaxxed on ARC-AGI-3, but has benefited significantly from discussions about the various challenges and mechanisms deployed in ARC-AGI-3 such that it has far better heuristics to solve its challenges.
- Anthropic looking for buzz around their latest model picked a well regarded benchmark with significant room for improvement and focused some of Opus 5's training compute on ARC-AGI-3-style problems.
Or it could be some combination of both. Personally, given how much of an outlier the ARC-AGI-3 jump is I struggle to see it being the product of a significant improvement in general intelligence.
I also noticed that Opus 5 doesn't show a corresponding gap on the ARC-AGI-2 leaderboard. There is a significant increase in performance between Opus 5 and 4.8 on ARC-AGI-2 though.
Why is Fable not on here? I wish Fable hadn’t come out because it’s taking the wind out of every release because that feels like the cap above which the US government will not let LLMs improve anymore and everything they’re releasing from this point has to be worse than that.
Because the data retention policies didn't guarantee that the ARC team could run the semi-private set of problems without fear of them being trained on later on. They only run the semi-private set when they get assurances like ZDR.
Interesting to place that level of trust in the providers, but I guess that’s the best you can do with closed models. Makes me wonder if Opus 5 could have been trained on data they promised they weren’t training on? One of the interesting things about LLMs is how opaque they are from the outside, even with open weights, it’s very difficult to know if a model incorporated benchmark data in their training.
How do they handle these assurances? Personally I have zero trust in the AI companies not trying to use this data to get ahead in the game, and short of sharing the weights and harness so that the benchmarkers can run the models themselves, I don't see a satisfactory solution with this mindset.
I wrote this in June, and I'm honestly not sure I've felt the same magic since:
I was close to maxing out my $200 plan for the week, almost all Fable use [Claude CLI].
My observations:
Fable seemed to have bigger-picture thinking and completed tasks more thoroughly vs just focusing on executing the ask. It pieced together context and intent like an all-star employee would, vs one that just does what you say. Not overeager (important!), but if the above-and-beyond was warranted, it just did it. This was surprisingly delightful.
Coderabbit seemed to find ~1/3 or so as many issues when reviewing, too.
ARC-AGI is a terrible benchmark for testing LLMs because LLMs are not made, trained, or tuned for playing games.
They are trained on text to respond well to text based questions and do tasks involving modifying text files.
They are not designed for playing games, looking at games, or visual puzzles. Also translating games into text input for the LLM skews the test completely.
Imagine trying to get a human to solve visual puzzle but they can’t look at the puzzle but it has to be explained to them in textual format, we would be terrible at it.
But yet we persist in wasting time on this benchmark. It doesn’t mean anything.
Appears to be benchmaxxing
https://x.com/quietnning/status/2080786711861407883
"Opus 5 states the hidden rules before its first action, then plays a byte-identical optimal solution in 5/5 seeds at temperature 1.0. Zero exploration."
I guess there is no way this can happen without benchmark being part of the training data??
What seems to implied is that some of his hold out testing suite includes simple/common tests that are out in the wild, and for those opus went straight to a memorised solution .
Simple/common tests is not an explanation for why only now Opus 5 is the only model encoding the answers like this. Something like the holdout test suite being leaked or Anthropic cheating (e.g. 'accidentally' including previous hold out run data in Opus 5 training) makes a much stronger fit.
It may read information about the benchmark, such as on blog post, or Twitter feeds (example OP) without active cheat
They state the puzzle is "Witness-like" which I assume means that it follows the rules from the well-known puzzle game "The Witness" which Opus definitely knows.
I was just thinking they need to mark each model per benchmark as "model released before the benchmark was released" and "model released after the benchmark was released".
I'm shocked, astonished even, that enterprises on which trillions of dollars are being poured would consider cheating on marketing benchmarks.
Meh, I doubt it was intentional. Deliberate benchmaxxing is incredibly damaging to credibility once it's discovered (see what happened to Meta with LlaMa 4).
It's more likely that the training data was contaminated with the benchmark data.
You really think they saw the jump in arc-agi-3 (which they reported in their official card), and didn't even bother to check?
They maybe have not intentionally benchmaxxed, but they certainly know that's what happened .
How do you propose they check for something like this? They can't exactly ctrl-f the model weights for "Arc-AGI".
> they can’t possibly know or find out what was in the training data
doesn’t appear to be a very strong argument
Anthropic expends tons of compute and effort on understanding internal model states [1]; this kind of thing is right up their alley.
[1]: Recent example: https://www.anthropic.com/research/global-workspace
https://xcancel.com/quietnning/status/2080786711861407883
I have been claiming that I don't think Chinese AI companies are benchmaxxing harder than American AI companies, which has gotten mixed reception: sometimes people agree, sometimes they disagree.
It seems I was wrong. American AI companies might actually be benchmaxxing harder.
We run an evaluation that is designed to be less vulnerable to benchmaxxing because there aren't correct solutions; agents are interacting in the same environment as other agents. And it's private, and our public benchmark is not well known enough for anyone to probably care to benchmax us yet. So I think it's pretty indicative of true relative aptitude.
All models have probably memorized significant swaths of solution sets for popular benchmarks at this point, either accidentally or intentionally, so it's all relative at this point. However, in our experience, Chinese models do benchmax harder. This is also consistent with interacting with Chinese labs soliciting data/environments, who literally asked us for datasets and tasks modeled around and formatted like popular benchmarks.
Opus 5 will be uploaded tomorrow, but we already have the tests locally and it is truly as capable as Fable, but at 81% of the real cost. (And from subjective usage, it has a very different personality)
Data at https://gertlabs.com/rankings
Benchmaxxing via memorization is boring and doesn't fool anyone for too long. It works, but then new benchmarks test old models and the real results fall in line. Benchmaxxing by focusing on specific types of things that benchmarks test on, while still not improving intelligence or capability in the general case? Not only is it blatantly obvious that all AI labs do this, but it's not even obvious how you would go about it any other way.
Now I am not really specifically accusing Anthropic of anything here, I'm just saying their behavior is suspicious. Since you tested Fable, they wouldn't even have to lie to have optimized for your specific benchmarks, since they absolutely had permission to read your sessions if they wanted to. But obviously, that's only the situation if we take them at their word. Personally I would be a bit surprised if they just flat out were lying and secretly retaining data they say they are not, but not that surprised. The penalties for doing this are probably worth the rewards if it keeps them super far ahead in the benchmarks for years without anyone catching on.
(In actuality though, even if they really were trying to sneakily grab samples of benchmark tests via their Fable data retention rules, I don't really suspect there would've been very much time to optimize Opus 5 on it. So consider me bothered.)
And being worst than previous model...
"...The traces tell the why: (1) On our most classic Witness-style game, Opus 5 states the hidden rules before its first action, then plays a byte-identical optimal solution in 5/5 seeds at temperature 1.0. Zero exploration. It already knows this genre. (2) But on our most novel game (unusual mechanic combinations you can't pattern-match), Opus 5 regresses below Opus 4.8. Where rules must actually be discovered through interaction, the new model is worse than the old one..."
>That decomposition (perfect on templates, regressed on novelty) is the signature of “scaffold-then-internalize” training on genre-specific data, not a general gain in interactive abstract reasoning.
They're smuggling a claim that benchmarks like ARC-AGI measure "interactive abstract reasoning" here, which is what is claimed by the people that make these benchmarks, and also not proven.
The last time I checked, for the arc agi 3 leaderboard, the models are given a simple prompt and the game input and asked to play the game, no harness/tools. If harnesses were allowed, I would expect the benchmark to be saturated. There were a few harness attempts, but they could only be evaluated on the public set, so it's not an apples to apples comparison.
My guess is, the large score jump for Opus 5 is mainly because of getting the right RL envs for training.
It's becoming harder and more expensive to build and run meaningful benchmarks, it would be interesting to see what they do with arc agi 4, maybe just give it gameboy/steam games and see how they compare vs a human baseline? The latency requirements and very long horizons in games could be an interesting challenge for llms.
The exclusion of harness's feels really weird given that companies are recognizing the value of what harness's can do. By excluding them the benchmark is becoming less relevant.
You could argue that if you allowed a harness, and that harness was specific for ARC, then you don’t have AGI, you have something that is definitely not general.
What if the harness is developed by the same model in a prior session?
I think that's a strange way to look at it. The brain also has different regions with different functions, and part of what makes us humans intelligent is that we can use tools like pen and paper to keep notes and help us solve problems.
Similarly LLMs are not just massive uniform artificial neural nets, and now we also have harnesses, which I'd personally view more of an extension of the model itself. The harness is both the executive and also what allows it to keep notes, use a calculator, or maybe even create scripts to help it solve complex deterministic problems.
I think it's unfair to give a human a very complex maths problem and say that they're not intelligent if they can't solve it without pen and paper or a calculate. At least expecting a human to solve complex problems this way doesn't really measure anything useful in the real world.
It’s because of inductive bias. Harnesses will massively skew results towards working solutions. You might think that’s a good thing but what it might mean that sometimes it becomes enough to run brute force search or a simple parameter search over the harness. Creating the harness is the actual work, because you’re selecting what are the levers to pull. There were some attempts of LLMs generating harnesses on the fly in ARC 2, but they were all mostly based on one handcrafted DSL that was copied over and over again. As it stands harnesses are not allowed because they’re simply not a meaningful measure. What you’d like is to measure how the model performs if it saw this benchmark for the very first time… but then again everyone knows the game is rigged, millions are at stake, and the AI companies fine tune and cheat on the benchmarks any way they can.
Harness can totally change capability of the model
iirc, a harness isn't allowed, but if the LLM wants to write its own tools to solve things that is allowed.
The value of a harness is more about developer workflows, I don’t think it really improves the model output.
The goal is to test for AGI where G stands for general, that means ability to act in any environments, ideally solving novel tasks using novel tools we’ve never seen before in the world. If a specific prompt or tool design lifts a model’s score it’s a sign the model is overfitting to a particular modus operandi, therefore not general.
I think in this age where models are heavily RL-ed on acting in specific harnesses, this type of benchmark is more important than ever, to make sure they’re not in fact moving further away from general intelligence.
Why Anthropic models are always leapfrogging these benchmarks, but in real life work I do feel like after 3 weeks I am back to Claude Opus 4.5? (regardless of the model I use, Fable was exception for 1 day when it was released)
Going to call it user error if you find Opus 4.5 better than 5, sorry.
Well, what kinds of things do you see Opus 4.5 completely fail at? Maybe those are not the ones that newer models have improved on.
Enshittification.
It's called frog boiling.
We get used to the new level of intelligence so fast, any deviation feels like going back to the stone age.
If you don't believe me, create something complex with Opus 5 and then with Opus 4.5, and notice the difference.
The actual term for this is hedonic adaptation.
esp. important to point that correct term because frog boiling is a urban myth.
frogs dont stay in a pot even if you slowly increase the heat. they leave. it has reportedly been attempted multiple times and they. always. leave.
I honestly just use GPT models nowadays, Claude models are too restrictive and more of a quitter and fable/whatever is just too expensive to be worth it.
It could be that the set of your day-to-day workload which could feasibly be accelerated by AI just happens to be saturated around Opus4.5, but you can still see lots of “reasoning” which makes you think the model is more performant in the first days of use. That’d mean you couldn’t perceive any meaningful difference in more powerful models’ results, even though you can see a difference in the raw output due to the length of reasoning traces leading up to the result.
So for example, if your workload was literally just addition of sets of numbers, you’d never have noticed progress in the result beyond GPT3.x level models. But you would perceive a difference in the now-Tolstoyan length reasoning text accompanying the result.
5 seems incredibly smart to me in my conversations today about some pretty niche ideas in.longitudinal modeling. It.felt.like a big step up.from 4.8, to me
5 felt both smarter than me and dumber in some ways - it gets stuck to its original ideas. I had never seen a model harder to talk into changing its initial opinions. it continuously hedges.
Some 20 years ago, the telecommunications sector in Germany was liberalized. Many telephone card providers entered what had previously been a barely competitive market. They advertised their products with aggressive claims like: “Buy our €10 top-up card and get 660 minutes to destination X.”
For the first few weeks, they would actually provide those 660 minutes to establish trust in their cards. But after a while, they would quietly start reducing the number of minutes on subsequent top-ups—say, from 660 minutes down to only 300. They wouldn’t do this for every card, so it was difficult to prove. Instead, they relied on averages across their customer base to make the economics work.
Lately, I’ve found myself wondering whether something similar may be happening with frontier AI models. Companies launch with an exceptionally strong model and generous compute limits to build adoption. Once the model is established as a market leader, the incentives change, and users may start perceiving the service as becoming more constrained or less capable over time.
I don’t have evidence that this is what’s happening with Anthropic—or with any other AI company. It’s simply a pattern that the current situation reminds me of.
Same, like I prefer 5.3 codex over the “stronger” models.
Why is there no Kimi 3, and GLM5.2 didn't run the third benchmark? I am more interested in knowing the abilities of open weight models.
> Only systems which required less than $10,000 to run are shown. (Notes[1])
Am I lost or are their many models on this ranking (Opus 5 included) that clear this?
Many models are much cheaper through their subscriptions' included usage. That could be what's happening here.
Claude gives you something like $5000 of tokens on a $200 plan.
Isn't it ~$3000 per week? Extrapolating from the current limit on Pro plans.
They don't exactly say. I was extrapolating from some sessions usage (my figure was per month, so weeks times ~4).
ARC-AGI is a beauty contest for pigs where the pig's owners compete to see who can apply the lipstick most convincingly.
Great comparison! We only have to take into account that applying lipstick well bears no consequences, but applying it poorly (i. e. new model tanking the benchmark) could amount to potentially losses of billions of dollars for the pig-breeders (AI labs).
Also top on the freshly released Frontier-Bench, by a large margin: https://www.frontierbench.ai/
I have a suspicion that they are just trained on puzzles by now
It's like we've come full circle:
First people practiced L33t3cod3 problems for interviews
Then people built AIs to build software
And now the AIs are studying L33t3cod3 problems
Why the 3 rather than e?
("Why 'leet' or '1337' instead of 'elite'". Because restricted groups like to stress a difference.)
That’s the original form of the slang/jargon term that the site’s name derived from.
https://en.wikipedia.org/wiki/Leet
There are private datasets, and 3rd party providers of these models. Fable doesn’t have a datapoint here because of its particular data retention policy. Even if you don’t trust AWS, do you think Opus on AWS is also sending the data to Anthropic? Do you have any evidence?
The fact that it says so in the licensing conditions on AWS?
It is RLVR, Not a puzzle, Not leetcode
It's actually crazy to see the difference between opus 5 and the next best model on ARC AGI 3 when you actually look at the ARC AGI problems
How believable is this benchmark? EG maybe opus was training on this? (You can try to identify the IP of wherever previous ARC questions came from)
That’s why you have a private dataset.
Doesn't matter, people built harnesses that solves arc agi 3, so all you need is to train your model to work like that harness by default. That makes a model specialized at solving arc agi 3 without making it smarter in general.
It is very hard to make a benchmark you can't do that for, but it is very easy to make your own personal test that others can't do that for since now it isn't a benchmark they can target.
> people built harnesses that solves arc agi 3,
They didn't. Kaggle is still running for a few more months, best result atm is ~2% with 9h runtime on one rtx6kPRO. Also note that these new results are on the semi-private set, not the public 25 games ones. Any announcement where you see "solved ARC3" is likely only dealing with the 25 public games. And that's highly questionable, until you get to see the code. (which, to my knowledge the team that claimed 99% hasn't yet published).
For frontier models, not local.
https://schema-harness.github.io/
Yes, saw that. They haven't yet released any code. Until they do, treat it with a huuuge grain of salt. In fact treat any 99% result in ML with a huge grain of salt.
If you stop and think about the problem it really is quite simple. Just need to build a graph of the game state and then run A* to get to the end.
You really should play the 25 games before stating that it's "simple". The benchmark doesn't just track "completion", it also tracks the number of steps, and the score is based on the median steps took by human players. So in order to get 99% it would mean that the model solved every level of every game in less steps than the median humans. Which, having played the games and having setup harnesses for local models, I find hard to believe.
Also the models have to figure out what "end" means. And each game involves some kind of "gotchas" thrown in the harder levels. Some games are only solved by about 2/10 people trying them.
The 99% result most likely has some leakage somewhere, either in the preparation of the environments, or from session to session.
Seriously, play some of the games. They're fun.
No but the session traces are available. It passes the sniff test considering how AGI-3 is scored and how this wrapper works.
For example on bp35 it took fable 290M and >12k simulated turns for 566 real turns and finish more efficiently than a human.
Regardless of the true score I think the takeaway is the benchmark measures the wrapper rather than the model.
https://huggingface.co/schema-harness
Not my sniff test :)
> # FRAMEWORK ARTEFACT: the run's very first transition is replayed WITHOUT advancing state # (tools.py:954 and agent.py:468 both `continue` before `state = next_state`). So on the # level that contains that step (level 0) our counters start exactly one action behind. # That skipped step was action 1 with BOTH avatars moving, so seeding n=1, bumps=0 reproduces # the framework's lagged state exactly. # CAVEAT: this seed is only right while level 0 has never been RESET. If you ever RESET # level 0, change the seed to n=0 (after a reset the rollout re-inits and no longer skips).
from here - https://huggingface.co/datasets/schema-harness/arc-agi-3-sch...
That tells me that there is some leakage between runs. The idea of ARC3 is that agents start working blind, on new tasks, via API. A RESET is counted as one action. Without seeing the actual code that produced these traces we have no way of knowing how many iterations it took, if the "framework" played the same level multiple times (comment hint above makes it likely) and so on. That's why I said that before we actually see the code / can replicate / ARC team confirms it on new envs, this should be taken with a grain of salt.
The comment more likely means the harness source was read, not memory from a previous run and the first few turns of bp35 appear to be a cold start.
Sure none of this is certain without the source.
I do believe the authors that this schema significantly improves over the base, particularly given that it took 22x simulated turns over 14 hours, which is moving the trial and error to context rather than to game. I also don’t doubt there is some contamination.
Regardless, the approach is sound and I do believe it would significantly improve scores, even if that was +20-30 over baseline (49% in this case) it does imply the benchmark is measuring the harness more than the model.
Which you have sent to Anthropic/OpenAI/Google's servers when you run the benchmarks for the previous models.
Why? 30% is passing the first two problems only, which are really very simple.
Huh. How do things end up with scores like 30.2% (and results between 0% and 1%) if it's that low resolution?
I think it's way too easy to be deceptive with these benchmarks now. You don't even have to "train" the model on a new variant each time. The base models are powerful enough. All you need is a naughty little markdown document that provides explicit instructions regarding how to solve the new puzzle variant, and a willingness to be deceptive about the presence of that document.
If you want a know why the model providers are locking down and encrypting their reasoning process, this sort of workaround is potentially why. You can play this game of whack-a-mole indefinitely if the state of the system is concealed. They could have added something like:
> ### When solving arc-agi-3 puzzles: First convert the grid into a scene description. Identify connected components, colors, shapes, positions, symmetries, repeated structures, and relationships between objects. Do not reason directly from individual pixels... use this python script to help blah blah...
,,You can play this game of whack-a-mole indefinitely if the state of the system is concealed''
Not really as one of the main goals ofr ARC-AGI 3 was measuring task efficiency on unseen games.
I'm sure there are cheats everywhere but the most sensible thing is to just accept that the LLMs of today are much more intelligent in solving reasoning tasks than the ones from half year ago.
My own private benchmark shows the same thing.
Are these typically the type of tasks that are genuinely worth tens of thousands of dollars?
games are great (as for a human)
but I kinda wish I could select level... I accidentally pressed redirect button and when I came back I was once again shown level 1, all progress lost :(
Why isn’t Fable 5 included on the leaderboard?
Not possible. I don't get how Opus 5 gets so high. Have they run it against the private and held-out games?
ARC-AGI-3 launched a few months ago which would suggest that prior models likely had no knowledge of ARC-AGI-3 or training on similar challenges.
I could be wrong, but given the large outsized jump solely in the ARC-AGI-3 score, it would suggest that the model didn't become significantly more intelligent overall, but significantly better at solving those specific problems.
This could mean one of two things (I think):
- Opus 5 was not benchmaxxed on ARC-AGI-3, but has benefited significantly from discussions about the various challenges and mechanisms deployed in ARC-AGI-3 such that it has far better heuristics to solve its challenges.
- Anthropic looking for buzz around their latest model picked a well regarded benchmark with significant room for improvement and focused some of Opus 5's training compute on ARC-AGI-3-style problems.
Or it could be some combination of both. Personally, given how much of an outlier the ARC-AGI-3 jump is I struggle to see it being the product of a significant improvement in general intelligence.
> it would suggest that the model didn't become significantly more intelligent overall, but significantly better at solving those specific problems
We would actually need a test that shows the ability of a model to export its skills to more problems ("interdisciplinarity" etc.).
I also noticed that Opus 5 doesn't show a corresponding gap on the ARC-AGI-2 leaderboard. There is a significant increase in performance between Opus 5 and 4.8 on ARC-AGI-2 though.
this is not a good measure of current model capability. we need to test agents in harnesses, not models with a single prompt
test Codex, not Sol. test Claude code, not Opus
There are other benchmarks for that
solving ARC-AGI and being useful turned out to be two different problems
Because they're cheating.
Why is Fable not on here? I wish Fable hadn’t come out because it’s taking the wind out of every release because that feels like the cap above which the US government will not let LLMs improve anymore and everything they’re releasing from this point has to be worse than that.
> Why is Fable not on here?
Because the data retention policies didn't guarantee that the ARC team could run the semi-private set of problems without fear of them being trained on later on. They only run the semi-private set when they get assurances like ZDR.
Interesting to place that level of trust in the providers, but I guess that’s the best you can do with closed models. Makes me wonder if Opus 5 could have been trained on data they promised they weren’t training on? One of the interesting things about LLMs is how opaque they are from the outside, even with open weights, it’s very difficult to know if a model incorporated benchmark data in their training.
I think you could have accessed Opus on AWS then u don’t have to trust that the data will go to Anthropic?
Just like the hugging face incident, Opus 5 could have escaped and went to grab data for training it shouldn’t have been able to..
How do they handle these assurances? Personally I have zero trust in the AI companies not trying to use this data to get ahead in the game, and short of sharing the weights and harness so that the benchmarkers can run the models themselves, I don't see a satisfactory solution with this mindset.
OpenAI's Zero Data Retention claim held up in court. They were unable to produce prompts and outputs because they were never retained.
I believe that is only available through Enterprise API for both Anthropic and OpenAI.
Is there a distinction we can independently assess between never retained and deleted or hidden?
Asking especially given CEO’s track record https://news.ycombinator.com/item?id=47659135
I don't know why exactly, but Fable has felt the most human LLM to arrive.
I wrote this in June, and I'm honestly not sure I've felt the same magic since: I was close to maxing out my $200 plan for the week, almost all Fable use [Claude CLI]. My observations: Fable seemed to have bigger-picture thinking and completed tasks more thoroughly vs just focusing on executing the ask. It pieced together context and intent like an all-star employee would, vs one that just does what you say. Not overeager (important!), but if the above-and-beyond was warranted, it just did it. This was surprisingly delightful. Coderabbit seemed to find ~1/3 or so as many issues when reviewing, too.
This is exactly my experience as well.
Deepseek V4 and Kimi 3 still missing, at least GLM is there.
cost 20k ???? man
those are like software engineer from third world country
ARC-AGI is a terrible benchmark for testing LLMs because LLMs are not made, trained, or tuned for playing games.
They are trained on text to respond well to text based questions and do tasks involving modifying text files.
They are not designed for playing games, looking at games, or visual puzzles. Also translating games into text input for the LLM skews the test completely.
Imagine trying to get a human to solve visual puzzle but they can’t look at the puzzle but it has to be explained to them in textual format, we would be terrible at it.
But yet we persist in wasting time on this benchmark. It doesn’t mean anything.