METR audit says Anthropic’s model progress roughly doubled, largely from in-house AI

METR audited Anthropic for eleven weeks on company hardware (12 Aug–29 Oct 2026, 2.5 FTE, unpaid), finding model gains roughly doubled with ~80% from in-house AI research while Anthropic redacted key proof. Reviewers accessed unreleased benchmarks, training metrics, surveys, analyses and ~15 intervi

Hírek

METR audit says Anthropic’s model progress roughly doubled, largely from in-house AI