Independent evaluator METR reported that Anthropic’s scores on an internal benchmark had accelerated, estimating AI tooling accounted for much of the increase, while noting substantial uncertainty.
On Nov. 5, 2026, independent evaluator METR reported that Anthropic’s internal AECI benchmark scores had risen 1.9 times faster since October 2025 than in the period from March 2024 to October 2025. The estimated rate increased from about 15 to 29 index points a year; METR’s 95% confidence interval for the acceleration is 1.5–2.8x. It judged that AI-assisted research and development likely explains most of the acceleration, while warning that its broader estimate—roughly twice the previous pace—could plausibly be 1.3x or 3x. METR’s assessment drew on confidential data and staff interviews; Anthropic redacted much of the evidence the evaluator considered most informative, limiting independent scrutiny. METR said faster growth in training compute did not explain the change, but cautioned that benchmark gains may not translate directly to real-world capability. The finding matters beyond Anthropic: if AI tools speed model development, competitors and safety evaluators may face a faster-moving cycle.