METR: Anthropic Nearly Crossed The AI Acceleration Threshold

The METR team is an independent group of AI safety researchers. For 11 weeks, from August 12 to October 29, they reviewed part of Anthropic's public safety report. Anthropic's own safety policy requires it to disclose when AI agents speed up its research progress by more than 2x. In its report, Anthropic states that the acceleration factor is "less than 2." We checked their internal metric ourselves. It comes out to 1.97. That is just below the threshold.

Anthropic has full control over which metrics go into this acceleration factor. It also controls which sections of the report get cut before publication. So when we look at the key numbers in their report, we are mostly just taking their word for it. And the real threshold sits that close to being crossed.