How much is AI helping researchers build more capable AI? METR’s independent assessment published November 5 examines Anthropic’s internal data and staff interviews to explore that question.
What does 1.9 times measure?
METR estimates that benchmark scores on Anthropic’s AECI measure have grown at 1.9 times their earlier rate since October 2025, with a 95% confidence interval of 1.5–2.8. This is a change in the rate of measured benchmark progress, not a claim that every task is completed 1.9 times faster.
Looking more broadly at well-specified tasks, the researchers judge that model capabilities are probably improving at roughly twice the historical rate. They also say a rate closer to 1.3 or 3 times would not surprise them.
AI’s contribution and the evidence gap
METR considers AI use in research a likely major contributor and offers roughly 80% as a highly uncertain judgment. Interviews and staff self-reports informed the assessment; those reports are not equivalent to a controlled productivity measurement.
Anthropic redacted much of the evidence METR found most informative, citing confidentiality. METR says this makes it harder for the public to evaluate the findings independently. The figures are therefore estimates with meaningful limits, not a settled law of progress.
