METR Says Anthropic Is Improving Models About Twice as Fast, Mostly Via AI for R&D
Since August, a team within METR has worked with internal Anthropic information to assess whether AI itself is accelerating AI development. We are now publishing a report summarizing the results of this risk assessment.
Context
An outside team with deep access says AI is already speeding Anthropic's own model progress, which is the clearest public signal yet of recursive speedup risk.
METR (@METR_Evals) is an AI evaluation nonprofit. From August 12 to October 29, 2026 its staff ran a high access review at Anthropic on whether AI is speeding AI R&D. Anthropic could redact nonpublic details. METR also published a separate review of the automated R&D section of Anthropic's August 2026 Risk Report.
Deep dive
METR (@METR_Evals) posted on Nov 5 that a team inside the group had worked since August with internal Anthropic information to judge whether AI is accelerating AI development, and that it was publishing a report on that risk assessment.
In the public report, METR says that as of October 2026 Anthropic is likely improving model capabilities on well specified tasks at about twice the rate seen between March 2024 and October 2025, with wide uncertainty (roughly 1.3x to 3x). Benchmark scores on Anthropic's AECI index grew at an estimated 1.9x the earlier rate since October 2025. METR says it is confident the speedup was not from faster growth in training compute, and guesses that a majority (perhaps about 80%) of the current acceleration comes from AI used for R&D, though with substantial uncertainty.
METR also says it did not see clear bottlenecks that would block further acceleration over the next year, while warning that many of the most informative evidence lines were redacted by Anthropic. Staff interviews and surveys pointed to large uplift from AI tools, mostly as assistance rather than fully autonomous R&D, with some signs that more autonomous work could grow soon.
The lead post passed about 1.09 million views. Because Anthropic redacted much of the supporting evidence, the public has to lean on METR's judgment more than usual, which METR itself flags as a limit.
What to expect next
Watch for Anthropic's own promised analysis of its internal progress speed, for other labs to invite similar outside reviews, and for whether later public metrics match METR's about 2x claim. Policymakers and investors will ask whether this level of AI driven R&D speedup triggers RSP style thresholds. Follow up work that can share more unredacted evidence would matter more than another headline number.