An Outside Evaluator Says Anthropic's AI Is Improving Twice as Fast as Before
METR, an independent AI evaluation organization, reported on Nov. 5, 2026, that benchmark scores for Anthropic's models are rising at an estimated 1.9 times the rate they rose before October 2025, with a range of 1.5 to 2.8 times at 95% confidence. METR staff took that measurement inside the company from Aug. 12 to Oct. 29, 2026, with access to Anthropic's internal training data, scores for models it has not released, and interviews with its staff.
METR's report says the point of the exercise was to find out whether AI has begun speeding up AI development itself, a process the organization says could lead to the sudden emergence of extreme and risky capabilities. Anthropic names that acceleration as a key risk indicator in its own Responsible Scaling Policy. METR says it took no payment for the work.
In plainer terms, the report puts the shift at roughly two years of historical progress made over the past year. It then hedges wider than its own statistical range, saying it would not be surprised if the true acceleration were closer to 1.3 or 3 times the historical rate. The measure is AECI, Anthropic's own version of the public Epoch Capabilities Index, so the index being tracked belongs to the company under examination.
On what drove the change, METR's report states: "We are confident that the acceleration was not due to an acceleration in the scaling of training compute." It guesses instead that a majority of the current acceleration, perhaps around 80%, comes from Anthropic's extensive use of AI to do research and development, and calls that judgment substantially uncertain.
How far anyone outside can check that is limited, and the report says so. Anthropic held unilateral rights to redact non-public information, and METR writes: "While we were able to freely share our highest-level conclusions about risk, Anthropic redacted a majority of the evidence we found most informative for these conclusions." The public, in the report's own account, has to rely on METR's judgment much more than would be ideal.
Anthropic also cut part of the original scope midway through: a verification of how it splits computing power between capabilities work and safety work. METR says that component was canceled before it reached any substantive conclusions, and that Anthropic wanted to keep the engagement focused on risk assessment for now.
