METR assessed AI’s role in accelerating Anthropic’s R&D from August to October 2026. Anthropic’s model benchmark scores grew at 1.9x their prior rate since October 2025. This acceleration is not due to increased training compute. Interviews and surveys suggest AI use drives roughly a doubling in progress pace. METR estimates about 80% of current acceleration comes from AI. Anthropic redacted much of the most informative evidence. The public must rely on METR’s judgment due to these redactions. No clear bottlenecks were found to prevent further acceleration. The assessment covers internal metrics, benchmarks, and staff interviews. Findings include uncertainty about sustained acceleration and potential underestimation of risks.
What changed
Anthropic's model capability growth followed a steady historical rate before October 2025.
What it unlocks
Faster iteration on AI model capabilities through AI-assisted R&D processes.
- 1.9x acceleration in AECI scores since Oct 2025
- 95% confidence interval: 1.5–2.8x
- ≈80% of current acceleration attributed to AI
- 2 years of historical progress in 1 year
Sources