METR estimates Anthropic is improving its models twice as fast
An outside assessment with access to internal data puts the acceleration at about 2x on well-specified tasks. Anthropic redacted most of the evidence METR found most informative.
In the Daily Brief of 6 November 2026
METR, a non-profit that evaluates frontier AI systems, spent from 12 August to 29 October inside Anthropic’s research data: metrics on pre-training and reinforcement learning, training compute, internal benchmark scores and staff interviews. Its report, published on 5 November, says Anthropic is likely improving its models on well-specified tasks at around twice the rate of March 2024 to October 2025. METR would not be surprised by a true figure closer to 1.3x or 3x.
The main measured result is that benchmark scores on AECI, Anthropic’s version of the Epoch Capabilities Index, have grown since October 2025 at about 1.9 times the earlier rate, with a 95% confidence interval of 1.5 to 2.8. METR says it is confident that faster scaling of training compute does not explain this. It attributes a majority of the current acceleration to Anthropic’s use of AI in research, mainly as assistance, and says it saw no evidence of significant bottlenecks over the next 12 months, while warning that its view of those bottlenecks changed several times.
The caveats are large. Anthropic redacted a majority of the evidence METR considered most informative, citing confidential information that could help rivals, so readers must rely on METR’s judgement. METR also says the index may overstate progress because it favours verifiable tasks. The findings arrive as Dario Amodei, Anthropic’s chief executive, has called for the pace of capability gains to slow.