METR finds Anthropic may have crossed its own AI acceleration warning threshold back in July

Why it matters

METR, a nonprofit that studies whether AI developers could lose control of their systems, gained unusual access to Anthropic's internal research data, including benchmark results from unreleased models, training metrics, employee surveys, and interviews with researchers. It concluded that the pace of improvement in Anthropic's models has roughly doubled since October, with most of that acceleration attributed to AI assisting Anthropic's own research. Anthropic's Responsible Scaling Policy treats exactly this kind of acceleration as a warning trigger, and METR suggests the threshold may already have been crossed. The audit's force is limited, however: Anthropic redacted much of the evidence METR considered most informative and canceled a planned review of how it splits compute between capability and safety work. METR still rates acute catastrophic risk as low but sees a significant chance the acceleration continues, and internal dissent plus wide uncertainty bands complicate the picture.