What happened
From August 12 to October 29, 2026, METR staff performed a high-access evaluation at Anthropic (access to pretraining and RL metrics, training compute, internal benchmark scores, AI speedup analyses, model forecasts, and staff interviews/surveys) to assess whether AI is accelerating AI R&D. METR concludes Anthropic is likely improving model capabilities on well-specified tasks at around twice the rate of March 2024–October 2025 (high uncertainty; 1.3x–3x plausible), and that benchmark scores since October 2025 have grown at an estimated 1.9x the pre-October 2025 rate [95% CI 1.5–2.8x] — 'roughly two years of historical progress over the last year.' METR is confident the acceleration is not from scaling training compute. A majority of the acceleration is likely attributable to AI-assisted R&D, with staff self-reporting dramatic uplift. METR saw no evidence of significant bottlenecks or strong diminishing returns over the next 12 months, so further acceleration is a real risk; it notes Anthropic redacted a majority of the most informative evidence, so the public must rely heavily on METR's judgment.
Why it matters
This is the first high-access, independent quantification of whether frontier AI is self-accelerating AI R&D — a direct input to loss-of-control risk assessments and to whether lab safety frameworks (RSP thresholds) should treat recursive capability acceleration as a nearer-term exposure.
Action needed
Brief the board and risk committee on the implications for capability-acceleration assumptions; review whether internal AI-R&D speedups warrant earlier RSP/safety thresholds.