Outside review finds Anthropic improving models about twice as fast as a year ago

Anthropic is improving the capabilities of its models on well-specified tasks around twice as fast as it was between March 2024 and October 2025, according to an independent assessment METR published Thursday and reported by the Associated Press. METR estimated Anthropic's AECI benchmark scores have grown about 1.9 times faster since October 2025 than before, with a 95 percent confidence interval of 1.5 to 2.8 times. The nonprofit said a majority of the current acceleration is likely attributable to Anthropic's use of AI in its own research and development, not to faster scaling of training compute.

METR cautioned that the estimate carries high uncertainty and that it would not be surprised if the true figure were closer to 1.3 times or 3 times. It said it did not see clear bottlenecks that would stop further progress over the next 12 months, while noting its view of those bottlenecks changed as more detail arrived. Anthropic's Responsible Scaling Policy includes a 2x acceleration threshold; METR said Anthropic's earlier argument that it had not crossed that line was overstated, while most researchers at both organizations still leaned toward it not having been crossed by mid-July.