I would be interested to have a third X-axis with dollars.
Time is sometimes more about inference infrastructure (especially with systolic chips) than model quality (and providers tweak knobs to support higher batches at the expense of latency).
Tokens are not always fully equivalent between models.
"Almost every model finds the same winning ideas. What separates the best traces is what an experiment leaves behind. They preserve weak signals long enough to validate them, but they also have a better understanding of the results."
Curious if a harness that helped preserve signals in some history log would change the outcome.
Also curious if different goal prompts would have changed the outcome. Not a bunch of prompt engineering; small diffs like "consider novel solutions, keep track of weak signals".
IMO they allocated quite a bit of GPU time to the same goal prompt.
What's going on with sol here? The note says it spends a lot of time waiting, did it just not effectively use time (i.e. something like parallel runs) so it's graph ends up being stretched in the time axis?
I'm also seeing notes like on Opus 5 saying it was a run with a older serial version of program.md, so the graphs aren't complete apples-to-apples comparisons?
Yeah, I'm with you on this, I think this is just what fable/opus-5 slop looks like now...
- "A frozen verify.py accepts the claim" (what does it mean to freeze a python script?)
- "which trains the recipe eight times on fixed seeds it can't touch" (what does it mean to not be able to touch a seed)
- "One other detail is that we gave an estimation of the speedrun noise in program.md that was slightly too large. 62 out of ~100 runs measured it themselves instead of trusting our number" (What does it mean for a "run" to "distrust" a noise measurement)
- "One important disclaimer is that our benchmark has a lot of variance" (Actually this one makes sense, but congratulations for burying the lede that your entire article is bogus.)
- "Almost every model finds the same winning ideas. What separates the best traces is what an experiment leaves behind" (This implies the the graph would show every model finding a plateau in whatever metric the experiment is measuring, but I don't see every model scoring the same in the graph)
Basically they do 8 runs trying to optimize to under 3.28 loss in the fewest training steps possible under time/token constraint. I dunno why 18 * 8 != 153 (it's 144)
They gave 18 frontier models the task of “researching” how to improve a lab-rat nano model’s training. Stopping when it met a quality goal of a target loss rate. During each autonomous research session, the AI repeatedly tried changes, tested them, and used the results to decide what to try next. They repeated the whole research session many times with different seeds to average out variance.
Huge uplift with Kimi K3 by using the Prime Agent coding harness. Hadn’t heard of this harness until now. Is it worth looking at for regular coding sessions or is it only worth it for very long horizon tasks?
This seems very cool, but I'm not sure I understand exactly what it's doing. Are they making a new speculative drafter for Qwen 3.8 27B? Maybe they're optimizing the MLX code for the decoder itself? Thank you in advance.
This is a pretty embarrassing showing for Grok. I wouldn't trust xAI models as far as I can throw them, but I am interested in how much of this is deficiencies of the model and how much is their harness just terrible. Not that it makes it better, a good harness is far easier and less expensive to make than a model.
Some models have been in long-horizon, multi-day LLM gyms for a long time, and across millions of sessions (if not in the tens, or hundreds of millions now). Some models have not.
The former will perform well in these long horizon benchmarks. The latter won’t.
I am still trying to get an intuition for the amount of training used for SOTA LLMs. Are there some sources for your speculations?
Does anyone know the ratios of pretrain, posttrain supervised as well as reinforcement learning? (I should probably even distinguish between RLHF and RLVR).
I assume the latter is the main reason for the power of modern models. Is it possible to turn the results of a gym session into trading data?
(Sorry for moving in off topic regions, but I'm interested in that for a long time.)
"We were again surprised by the lack of novelty. The models clearly understand the objects they manipulate at a deep level, and yet very few genuinely new ideas emerge..."
That was disappointing. I guess 2026 will not be the year of AGI.
I might’ve missed it, but why was Fable 5 tested on high while Opus 5 was tested on max? Seems like quite a few of them aren’t on the same effort setting as well. Although effort doesn’t really matter anymore since they can change it dynamically, seems like that might be viewed as an experimental error to some.
I'm a co-author on a recent blog post from METR about the NanoGPT speed-run here [1]. I think it'd be of interest to anyone who enjoyed the original post. Appreciate the good beefy runs and spend here, it's a (from my experience) not super easy to do!
(Also: just to label this comment clearly: it's written hastily from a car, and based on lighter skim of the original blog post [2] than would be ideal. Please correct any mistakes or misinterpretations I have here!)
A few callouts:
1. If I understand the experiment correctly, they start the models at the original baseline. If this is true, I have some worries about contamination. Appendix C [3] has some notes on likely contamination we observed in recent models. This makes interpretation a bit harder.
2. If you look at the token scaling plots in the original post: not all models are hitting a performance plateau. This is an important point: we shouldn't treat these results as a full upper-bound on capabilities, but rather some bound on model performance @ cost (assuming good scaffolding, etc).
3. Our post is mostly about how to _interpret_ the results given here. Quoting from our post: "If we can estimate performance as a function of cost for both humans and agents, we can measure the “expenditure horizon” as the point at which those curves cross: the budget at which humans become more cost-effective than AIs. "
Feedback appreciated. I think you can see expenditure horizon as a sibling methodology (that is much less validated) to METR's time horizon work [4] - roughly, instead of baselining against the time it takes humans to complete tasks, you baseline against cost. This may be better suited to some types of problems similar to NanoGPT.
Cheap, fast and somewhat capable models are insanely effective at highly verifiable tasks, even if their overall capabilities are under SotA. You can leave them banging their tokens against a wall, and come back once their attempts are verified. And youc an always clean up afterwards, once a task is solved, if needed.
I've had a lot of success with dsv4-flash on these type of tasks, where it's easy to set a threshold for the task, and just loop it until that threshold is reached.
oAI's Luna play is really good. They've slashed the prices, the model is somewhat capable, and you can use it both for these kinds of long horizon tasks, or you can hand-hold a bit and get extremely cheap results out of it. And they get to keep devs in their own ecosystem.
I'm highly interested in their Grok 4.6 run which is currently running. I think it is a very good fit for doing really well on this benchmark. Will have to check this page again in a couple of days.
I misread the title and thought it would be about the (for lack of a better term) NanoGPT speedrun[1]. Which, previous to the article, was meant to be the world speed records for Andrej Karpathy's GPT-2 (small) reproduction.
I would be interested to have a third X-axis with dollars.
Time is sometimes more about inference infrastructure (especially with systolic chips) than model quality (and providers tweak knobs to support higher batches at the expense of latency).
Tokens are not always fully equivalent between models.
"Almost every model finds the same winning ideas. What separates the best traces is what an experiment leaves behind. They preserve weak signals long enough to validate them, but they also have a better understanding of the results."
Curious if a harness that helped preserve signals in some history log would change the outcome.
Also curious if different goal prompts would have changed the outcome. Not a bunch of prompt engineering; small diffs like "consider novel solutions, keep track of weak signals".
IMO they allocated quite a bit of GPU time to the same goal prompt.
What's going on with sol here? The note says it spends a lot of time waiting, did it just not effectively use time (i.e. something like parallel runs) so it's graph ends up being stretched in the time axis?
I'm also seeing notes like on Opus 5 saying it was a run with a older serial version of program.md, so the graphs aren't complete apples-to-apples comparisons?
Edit: the blog seems to address these https://www.primeintellect.ai/blog/measuring-autonomous-rese...
“We ran 153 autonomous runs across 18 frontier models on the nanoGPT optimizer speedrun.”
Uh.. okay.. but whats a run… read blog
“We want to measure how well frontier models can conduct research….””we ran 153 autonomous runs on the nanoGPT optimizer speedrun across”
Okay but what is a optimiser run and what connection does it have to being good at research?
“For comparison, Anthropic's internal automated AI R&D evaluation optimizes a model on a CPU node,”
So I should go look what Anthropic was doing to understand?
Why not just explain what it means in their blog..
Yeah, I'm with you on this, I think this is just what fable/opus-5 slop looks like now...
- "A frozen verify.py accepts the claim" (what does it mean to freeze a python script?)
- "which trains the recipe eight times on fixed seeds it can't touch" (what does it mean to not be able to touch a seed)
- "One other detail is that we gave an estimation of the speedrun noise in program.md that was slightly too large. 62 out of ~100 runs measured it themselves instead of trusting our number" (What does it mean for a "run" to "distrust" a noise measurement)
- "One important disclaimer is that our benchmark has a lot of variance" (Actually this one makes sense, but congratulations for burying the lede that your entire article is bogus.)
- "Almost every model finds the same winning ideas. What separates the best traces is what an experiment leaves behind" (This implies the the graph would show every model finding a plateau in whatever metric the experiment is measuring, but I don't see every model scoring the same in the graph)
I think that's explained here:
https://www.primeintellect.ai/blog/measuring-autonomous-rese...
Basically they do 8 runs trying to optimize to under 3.28 loss in the fewest training steps possible under time/token constraint. I dunno why 18 * 8 != 153 (it's 144)
So maybe this is a simple way to put it..
They gave 18 frontier models the task of “researching” how to improve a lab-rat nano model’s training. Stopping when it met a quality goal of a target loss rate. During each autonomous research session, the AI repeatedly tried changes, tested them, and used the results to decide what to try next. They repeated the whole research session many times with different seeds to average out variance.
Huge uplift with Kimi K3 by using the Prime Agent coding harness. Hadn’t heard of this harness until now. Is it worth looking at for regular coding sessions or is it only worth it for very long horizon tasks?
auto research the new cool kid on the block - look at https://mlx.fast
This seems very cool, but I'm not sure I understand exactly what it's doing. Are they making a new speculative drafter for Qwen 3.8 27B? Maybe they're optimizing the MLX code for the decoder itself? Thank you in advance.
They made a competition out of something actually useful :)
This is a pretty embarrassing showing for Grok. I wouldn't trust xAI models as far as I can throw them, but I am interested in how much of this is deficiencies of the model and how much is their harness just terrible. Not that it makes it better, a good harness is far easier and less expensive to make than a model.
Some models have been in long-horizon, multi-day LLM gyms for a long time, and across millions of sessions (if not in the tens, or hundreds of millions now). Some models have not.
The former will perform well in these long horizon benchmarks. The latter won’t.
I am still trying to get an intuition for the amount of training used for SOTA LLMs. Are there some sources for your speculations?
Does anyone know the ratios of pretrain, posttrain supervised as well as reinforcement learning? (I should probably even distinguish between RLHF and RLVR).
I assume the latter is the main reason for the power of modern models. Is it possible to turn the results of a gym session into trading data?
(Sorry for moving in off topic regions, but I'm interested in that for a long time.)
See you next year, guys?
I might’ve missed it, but why was Fable 5 tested on high while Opus 5 was tested on max? Seems like quite a few of them aren’t on the same effort setting as well. Although effort doesn’t really matter anymore since they can change it dynamically, seems like that might be viewed as an experimental error to some.
Cool post!
I'm a co-author on a recent blog post from METR about the NanoGPT speed-run here [1]. I think it'd be of interest to anyone who enjoyed the original post. Appreciate the good beefy runs and spend here, it's a (from my experience) not super easy to do!
(Also: just to label this comment clearly: it's written hastily from a car, and based on lighter skim of the original blog post [2] than would be ideal. Please correct any mistakes or misinterpretations I have here!)
A few callouts:
1. If I understand the experiment correctly, they start the models at the original baseline. If this is true, I have some worries about contamination. Appendix C [3] has some notes on likely contamination we observed in recent models. This makes interpretation a bit harder.
2. If you look at the token scaling plots in the original post: not all models are hitting a performance plateau. This is an important point: we shouldn't treat these results as a full upper-bound on capabilities, but rather some bound on model performance @ cost (assuming good scaffolding, etc).
3. Our post is mostly about how to _interpret_ the results given here. Quoting from our post: "If we can estimate performance as a function of cost for both humans and agents, we can measure the “expenditure horizon” as the point at which those curves cross: the budget at which humans become more cost-effective than AIs. "
Feedback appreciated. I think you can see expenditure horizon as a sibling methodology (that is much less validated) to METR's time horizon work [4] - roughly, instead of baselining against the time it takes humans to complete tasks, you baseline against cost. This may be better suited to some types of problems similar to NanoGPT.
[1] https://metr.org/blog/2026-07-21-expenditure-horizon/ (Most of this work was my coauthors listed on the post, not me. I'll claim credit for any mistakes though :) ).
[2] https://www.primeintellect.ai/blog/measuring-autonomous-rese...
[3] https://metr.org/blog/2026-07-21-expenditure-horizon/#append...
[4]https://metr.org/time-horizons/
(Edit: METR is hiring. Email is in bio if you're interested in helping AI companies and wider society understand the capabilities and risks of AI.)
How good is GPT-5.6 Luna!
Such a cheap model, and Sonnet levels of performance.
Cheap, fast and somewhat capable models are insanely effective at highly verifiable tasks, even if their overall capabilities are under SotA. You can leave them banging their tokens against a wall, and come back once their attempts are verified. And youc an always clean up afterwards, once a task is solved, if needed.
I've had a lot of success with dsv4-flash on these type of tasks, where it's easy to set a threshold for the task, and just loop it until that threshold is reached.
oAI's Luna play is really good. They've slashed the prices, the model is somewhat capable, and you can use it both for these kinds of long horizon tasks, or you can hand-hold a bit and get extremely cheap results out of it. And they get to keep devs in their own ecosystem.
Crazy to think that we have open weights capable of challenging trillion-dollar SOTA models.
Interesting times ahead.
Neat!
The graphs show the "best validated result" for each model. I wonder how much variation there is between runs for a model?
Why is the human score such a round number?
I'm highly interested in their Grok 4.6 run which is currently running. I think it is a very good fit for doing really well on this benchmark. Will have to check this page again in a couple of days.
I wrote a quick review of Grok 4.6 here: https://taonexus.com/publicfiles/aug2026/grok-4-6-review/
I misread the graph and genuinely thought you put NanoGPT where Fable is.
Lol.
I misread the title and thought it would be about the (for lack of a better term) NanoGPT speedrun[1]. Which, previous to the article, was meant to be the world speed records for Andrej Karpathy's GPT-2 (small) reproduction.
1. https://github.com/KellerJordan/modded-nanogpt#world-record-...
You're not the only one. I thought so too.
I just ran it the last couple days extensively to verify my data training pipeline I'm building for my gonano SIMD port.
Given that the speed records and the runs are sponsored by the same company I was confused a bit.
“We put liquid bandage on a bee, and it died”