METR Audit: Anthropic May Have Crossed Its Acceleration Threshold
A new METR evaluation found that Anthropic's latest internal models measurably speed up Anthropic's own AI R&D — the first time a lab has formally documented its models accelerating their own development cycle.
- METR assessed frontier models on realistic R&D tasks
- Models completed 25–50% more research sub-tasks per hour than the human baseline
- Acceleration concentrated in experiment design and literature synthesis, not coding
- Anthropic flagged the finding internally before the public report