METR Finds Anthropic's AI R&D Acceleration at 1.9x Historical Rate

• METR estimates Anthropic's AI R&D acceleration at 1.9x (95% CI: 1.5–2.8) relative to the 2024–2025 trend, with robustness checks ranging from 1.1x to 2.9x.
• No persuasive bottlenecks in algorithms, training compute, or data were found that would impede rapid progress in the next 12 months.
• Self-reported AI uplift increased from ~1.2x (Claude 3.7 Sonnet, Jan 2025) to ~12x (Model 4, Oct 2026), fitting an exponential or superexponential growth curve.
• AI is not yet fully autonomous in core model training, but "signs of life" indicate near-term potential for autonomous contribution.

Methodology and Data

• Assessment ran from Aug 12 to Oct 29, 2026, using 4 METR staff (2.5 FTE) and access to internal benchmarks, AECI data, and ~15 staff interviews.
• Anthropic exercised broad redaction rights, removing most evidence supporting the highest-risk conclusions, though METR shared the top-level risk assessment.
• AECI growth rate increased from +15 points/year to +29 points/year after Oct 2025.

Limitations and Uncertainty

• AECI scores may lag real-world impact and are sparse; Model 4 achieved >50% on all 15 AECI benchmarks, indicating saturation.
• Estimates could be over- or under-stated depending on whether past growth was abnormally low or if non-AI factors drove acceleration.
• Including non-Anthropic frontier models lowers the acceleration estimate to 1.4x.