Editor | Zenan

AI helps to accelerate the development of the next generation of AI, and specific numbers have finally emerged.

On November 5, the third-party assessment agency METR released a survey report on Anthropic, indicating that the benchmark capabilities progress made by the company in the past year is roughly equivalent to the progress made in the previous two years.

From August 12 to October 29, the agency spent about 11 weeks examining internal data from Anthropic, including model performance, training computing power, research efficiency, and interviewing researchers.

Their analysis utilized Anthropic's comprehensive capability metric AECI. Data shows that since October 2025, the progress rate of large models on this metric has been about 1.9 times faster than before, with a 95% confidence interval of 1.5 to 2.8 times.

Since October 7, 2025, this score has jumped from about 15 points per year to about 29 points per year, achieving in one year what previously took about two years, roughly equivalent to the transition from Claude 2.1 to Sonnet 4.5. Anthropic's own risk report had previously assumed that Mythos Preview was a one-time leap, expecting a return to the original trend afterward; however, new models have emerged, and the new slope continues.

Looking across the industry, the acceleration of leading models is only 1.4 times.

More critically, METR is quite confident that this acceleration does not stem from the growth rate of training AI computing power. The main reason for the improvement in AI capabilities comes from AI helping itself.

In the interviews, a recurring phenomenon was that researchers could finally experiment with more ideas.

In the past, an experiment might take one or two days to implement, with time and computing costs being insurmountable, leading to many ideas being shelved before they could be verified. Nowadays, researchers can hand multiple ideas over to an Agent to see which ones are worth pursuing further.

One researcher described that their work is now primarily limited by ideas and computing power, with coding no longer being a bottleneck. Some even feel that after using Claude, they have the support of a postdoc and graduate students to advance projects. Experiments and implementations have become cheaper, making researchers more willing to try, and many experiments that were once deemed unworthy of effort now have the opportunity to turn into new research directions.

METR's report also revealed some internal progress that was previously unseen externally. Their analysis included three models from Anthropic that have not been publicly disclosed, referred to as Model 2, Model 3, and Model 4. They completed training between July and October of this year.

The strongest currently is Model 4 (some speculate it is Claude 6), which has already scored over 90% on many common benchmarks. Anthropic employees internally estimated its impact on their work efficiency to be about 12 times in a vote held in October.

These models have not yet been released, and it is not guaranteed they will become public products, but their existence serves as a reminder that merely watching press releases and public rankings may not give a complete picture of what is happening inside major AI companies. From about 1 time in early 2025, to about 2.5 times for Opus 4.5, about 5 times for Mythos 5, and Model 4 jumping to 12 times, the exponential fitting suggests an increase of 3.9 times per year.

METR estimates that by October of this year, about 80% of the current acceleration in AI research is estimated to come from AI. The remaining 20% are non-AI factors, the details of which Anthropic does not allow to be disclosed.

So, is AI already able to independently develop its next generation?

Currently, there is still a gap to achieve complete recursive self-improvement (RSI) for AI to independently develop the next generation of models. Anthropic researchers believe that in the most critical part of model training, AI has not yet made highly significant and completely autonomous contributions; the main benefits currently still come from AI assisting human development.

However, Anthropic researchers have described a large number of autonomous contributions from AI as "signs of life," and have quantitative analyses documenting the proportion of tasks completed by AI versus humans in various stages, but the data was deleted by Anthropic.

As long as stronger models can help researchers develop the next generation faster, the feedback process could continue to enhance. How far the path of RSI can go depends on the unautomated components and whether resources like computing power can keep up.

Regarding predictions for the near future, METR does not foresee any significant bottlenecks that would impede rapid progress in the next 12 months but admits that they have changed their assessments multiple times as information increases, and their current confidence is limited.

Additionally, although Anthropic allowed the third-party agency METR to conduct an internal investigation, the current 100-page report has been significantly redacted by them. The most persuasive evidence has mostly been removed by Anthropic itself.

The accompanying review released on the same day also showcased the differences in risk statements between the two sides: Anthropic previously described a leading indicator as having a progress rate "of less than twice," while METR calculated it to be about 1.97 times based on its internal data, believing that a more accurate description should be "about twice." The twofold acceleration involves Anthropic's early warning threshold.

METR believes that the possibility that the relevant threshold was crossed as of July 15 cannot be ruled out, but most relevant researchers still lean towards the view that it had not been crossed at that time. Both sides also believe that the direct catastrophic risk associated with the relevant models at that time was low.

As the speed of model involvement in development increases, whether people can timely recognize changes and whether companies can adequately disclose uncertainties is becoming a pressing question to be answered before the next generation of AI arrives.

Reference content:

https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/#executive-summary

https://x.com/METR_Evals/status/2118412183860351084

© THE END

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到[email protected]，本平台相关工作人员将会进行核查。