METR’s assessment of Anthropic estimates that an internal capability index is rising 1.9 times as fast as the earlier trend. It considers AI an important contributor, with greater uncertainty about the causal share. Anthropic allowed publication of high-level conclusions while redacting much of the evidence the evaluators found most informative.
METR, an independent AI evaluation organization, has published an assessment of research acceleration at Anthropic. It estimates that growth in Anthropic’s internal capability index is 1.9 times the earlier trend, with a 95 per cent confidence interval of 1.5–2.8 times. METR considers AI an important contributor to the acceleration, while treating its causal share as more uncertain than the observed trend.[1], [2]
The assessment covered 11 weeks from August 12 to October 29, with approximately 15 interviews about internal experiments and research workflows. METR says it received no payment, while Anthropic supplied devices and free model use. Its quantitative analysis fits a curve to model scores, locating a breakpoint on October 7, 2025. The confidence interval resamples scores and does not cover every uncertainty about model selection or causation.[1]
Anthropic permitted high-level conclusions to be published but redacted much of the evidence METR found most informative, citing confidential information that could help competitors. METR identifies where those restrictions affect its conclusions. It also distinguishes assistance to human researchers from fully autonomous research, judging the latter not yet a highly significant contributor to the most important model-training work.[1]
