AI is now building AI: METR finds Anthropic developing models twice as fast
What's the deal? Anthropic is improving its AI models on well-specified tasks around twice as fast as it was a year earlier, according to an independent assessment published November 5 by the nonprofit Model Evaluation and Threat Research (METR). The group found that most of the acceleration comes from AI helping build AI, rather than from faster compute scaling.
The numbers: METR estimated a 1.9× speedup in Anthropic's rate of progress since October 2025, with a 95% confidence interval of 1.5× to 2.8×. That amounts to roughly two years of earlier progress compressed into one. Staff using Anthropic's best internal model self-reported being about 12 times as productive.
How they know: METR spent 11 weeks inside Anthropic, from August 12 to October 29, with what it called "unprecedented access to internal data" on pretraining, reinforcement learning, compute, and benchmark scores, plus staff interviews.
Why now? The report lands as Anthropic prepares for a Wall Street debut that could be the largest IPO ever. It also follows a September 12 essay in which chief executive officer Dario Amodei urged the industry to "slow the pace at which we improve the capabilities of AI models" so safety work can keep up.
What could go wrong? The findings carry heavy caveats. Anthropic redacted much of the evidence METR considered most informative, describing the cuts as "aimed at removing confidential information that could accelerate other AI developers." METR acknowledged internal disagreement and said it could be overstating results, since the acceleration may be limited to measurable, well-specified tasks.
What's the endgame? Anthropic has said its ClaudeDealroom has a profile for this one. Try Dealroom → models increasingly help develop newer versions of themselves — the company reports engineers now ship eight times as much code per quarter as they did from 2021 to 2025. Taken far enough, that points toward recursive self-improvement, where an AI system designs its own successor. Anthropic says it is "not there yet."
The signal: METR's estimate sharpens the tension Amodei flagged: AI-assisted progress may be outrunning the pace of safety work, though the report does not establish that safeguards have fallen behind. With details about AI risk still supplied mostly by the labs themselves, an outside assessment — even a redacted one — is a notable data point as frontier development speeds up.
Read more: metr.org, The Verge
Image credit: Department for Science, Innovation & Technology
