AIUpdated November 7
METR estimates Anthropic’s model capabilities improved nearly twice as fast
METR’s 11-week assessment drew on internal R&D data and staff interviews Anthropic made available. It estimates that by October, the rate of improvement in Anthropic’s model capabilities on METR’s benchmarks was about 1.9 times the rate from March 2024 to October 2025. METR tentatively attributes roughly 80 percent of the acceleration to AI use in research and development, but calls that a subjective, uncertain estimate. The uncertainty is substantial: Anthropic redacted most of the evidence METR found most informative, and much of it allowed multiple plausible interpretations. The assessment also did not find AI making fully independent, consequential contributions to the most important stages of model training.
Coverage started November 5
Key Actors
- Anthropic