METR Investigation Suggests Anthropic May Have Breached Its AI R&D Acceleration Warning Threshold in July

An audit by non-profit evaluation organization METR shows that the rate of progress on Anthropic's model development test scores has accelerated by roughly 1.9x since October 2025, likely crossing its internally defined 2x acceleration warning threshold in July 2026. METR estimates that roughly 80% of the speedup stems from AI-assisted internal R&D; however, Anthropic heavily redacted core evidence prior to the audit's publication and canceled a planned dedicated audit on compute allocation midway through.