Outside Evaluator METR Says Anthropic's AI Progress Has Nearly Doubled in Speed

The findings put a number on the risk Anthropic's own leadership has warned about most: recursive self-improvement, in which AI speeds up the work of building the next generation of AI. In a September 12 essay, chief executive Dario Amodei urged the industry to "slow the pace at which we improve the capabilities of AI models" and named METR as the kind of outside evaluator he wanted embedded in leading labs. Now that evaluator has found Anthropic's own pace speeding up, weeks before the company's planned stock market listing later this month.

According to Reuters, METR estimated that Anthropic's benchmark capabilities have improved about 1.9 times faster since October 2025 than before, and that AI accounted for perhaps 80% of that acceleration, a figure the group cautioned was uncertain. "We are confident that the acceleration was not due to an acceleration in the scaling of training compute," METR said. The Associated Press reported that METR found Anthropic's benchmark progress over the past year matched what it achieved between October 2023 and October 2025, and that the group saw no bottlenecks, such as limits on computing power or data, likely to hold progress back.

How the Review Worked

METR published a review of the "Risks from automated R&D" section of Anthropic's August 2026 Risk Report, alongside a companion assessment of AI research acceleration at the company. The work drew on about 11 weeks of confidential material, from mid-August to late October, including internal data, employee interviews and performance results from unreleased models.

According to the timeline in METR's review, Anthropic began sharing confidential information on August 12 and handed over the unredacted risk report two days later. METR sent its first draft on September 18, Anthropic gave final redaction guidance on October 28 and approved the redacted version on October 29. METR said where those redactions affected its conclusions, and the AP reported that they limited how precisely the group could state its estimates.

A Threshold Anthropic Set for Itself

Anthropic's Responsible Scaling Policy treats a doubling in the rate of AI progress, or a year's worth of advances arriving where two would normally be needed, as a trigger for automated R&D risk. In its public report the company rated that risk "Low" and said it did not believe its models met either of its criteria.

METR agreed that catastrophic risk from Anthropic's models as of July 15 was low, but went further than the company on acceleration. "We think it is plausible that they had crossed their early warning threshold at the time," the group wrote. It said the public version of Anthropic's report understated both the evidence of acceleration and the uncertainty around it. Where the public report described leading indicators as showing acceleration of "less than a factor of 2", METR calculated that Anthropic's private figures implied about 1.97, and its own measures ranged from 1.3 to 2.8 depending on definition and period.

The group listed reasons its findings could be too high. The speed-up may apply only to well-specified tasks that are easy to grade automatically, rather than messier research work, and Anthropic may have understated its earlier progress, which would make recent gains look larger. The AP reported that METR found AI not yet able to make highly significant, fully autonomous contributions to model training, although Anthropic researchers described "signs of life" that this could come soon.

Anthropic did not immediately respond to the AP's request for comment. METR said it could not rule out substantial further acceleration over the next six months, which it said could quickly push Anthropic's own risk ratings to "high".