METR says Anthropic’s AI models are improving at roughly twice the historical pace

Anthropic’s AI models are getting better at roughly double their historical rate, according to a new independent assessment from METR, an AI safety nonprofit. The report, released on November 5, 2026, puts a number on something the industry has speculated about for years.

The number is about 1.9 times faster. That works out to two years of progress packed into a single year.

What makes the finding notable is its explanation. METR attributes the bulk of the speedup to Anthropic using its own AI systems in research and development. In other words, the AI is helping build the AI.

What METR found

METR ran its evaluation over an 11-week window, from August 12 to October 29, 2026. During that period, the nonprofit had access to confidential data, interviewed Anthropic staff, and reviewed proprietary performance metrics.

The yardstick at the center of the report is a benchmark METR calls the AECI. By METR’s estimate, AECI scores have been climbing at about 1.9 times the historical improvement rate since October 2025.

The comparison baseline is the pace observed between March 2024 and October 2025. So this is not a measurement against some sleepy early era of AI. The old rate was already brisk, and METR says the new one nearly doubles it.

Roughly 80% of the acceleration, according to METR, traces back to Anthropic’s internal use of its AI systems for R&D. METR found the gains came mainly through assistive use, not fully autonomous processes. The models are speeding up human researchers rather than running the lab on their own.

METR also firmly ruled out one obvious suspect. Faster growth in training compute, the raw processing power used to train models, was not the driver here, per the report.

Anthropic’s response and the safety backdrop

Anthropic did not exactly push back. Co-founder Sam McCandlish expressed agreement with several of the report’s findings on social media.

The timing was also coordinated. METR’s report came out alongside Anthropic’s own Risk Report commentaries, suggesting the company is treating the findings as something to discuss openly rather than bury.

That openness has a clear motive in the subject matter. The report sharpened long-running concerns about recursive AI self-improvement, the scenario where AI systems improve themselves faster and faster in a loop.

METR’s own finding offers some partial reassurance on that front. Because the acceleration comes mostly from assistive use, humans remain in the loop.

What this means for the industry

There is also a measurement question to keep in view. The AECI is a single benchmark, and METR’s figure is an estimate drawn from an 11-week window. The 1.9 times figure is striking, but future assessments will show whether the pace holds, accelerates further, or settles back down.

The more durable takeaway may be the template itself. An outside safety group got access to confidential data and staff, and the company under review largely engaged with the results in public.