The independent nonprofit organization Model Evaluation and Threat Research (METR) from Berkeley, California, analyzed internal data from Anthropic and found that the company is improving its artificial intelligence models roughly twice as fast as a year earlier. The report was published on Thursday.
Anthropic gave the researchers access to internal data covering a period of about 11 weeks - from mid-August to the end of October - and allowed interviews with employees. According to METR, this was "unprecedented access" to the AI development process. The company is preparing for a stock market listing that is expected to potentially become the largest initial public offering (IPO) in history.
The researchers focused on two questions: whether the pace of growth in Anthropic's internal capabilities had accelerated, and what role AI itself played in that. METR estimated that the company improves models on tasks with clearly defined conditions roughly twice as fast as a year ago. At the same time, the group believes a significant part of the acceleration is explained by AI assistance: Anthropic previously reported that its chatbot Claude is increasingly involved in creating more advanced versions of itself.
The report contains important caveats. METR acknowledges that it may be overestimating the result: the acceleration may be limited to "well-specified tasks" - those that can be automatically checked or evaluated using a language model - and may not extend to "messy" tasks that are harder to formalize. In addition, it is possible that Anthropic understated earlier achievements, which would make the current figure look higher. Nevertheless, the researchers "confidently" ruled out growth in the computing power used to train models as the cause of the acceleration. Usually the volume of compute is directly linked to model performance, but here, according to METR, something else is at play.
The group also examined Anthropic's public risk report published in August, its non-public version and other non-public data. The conclusion: the public version "generally understates the level of acceleration taking place."
Separately, METR notes: the company made edits to the report that limited the group's ability to report more detailed estimates of the acceleration; within the METR team there were disagreements and uncertainty, but overall the organization believes the pace of improvement has doubled. "Anthropic appears to have made the same progress on tests in the past year as it did between October 2023 and October 2025," the report says.
The researchers warn of a potential risk: if AI begins to improve itself, the acceleration could compound. Of particular concern is that they found no "bottlenecks" - constraints on computing power or data - that would hold back Anthropic's progress. That said, according to METR, AI is not yet capable of making "highly significant and fully autonomous contributions" to the most important parts of the model training process. However, Anthropic employees described "signs of life" indicating that this could happen soon.
Curiously, it was precisely during the evaluation period that Anthropic head Dario Amodei called on the industry to slow down. In a September 12 essay, he urged the AI community to "slow down the rate at which we improve AI model capabilities." Amodei also spoke in favor of bringing external evaluators into leading AI labs and called METR a suitable independent evaluator.
METR is funded by donations and does not charge for its work, but the companies it collaborates with provide access and tokens - units of text processing used for evaluation, research and engineering. Anthropic did not promptly respond to a request for comment.