METR finds Anthropic model capabilities improving nearly twice as fast with AI help
Independent high-access assessment: AECI benchmark growth ~1.9× since Oct 2025; majority of current acceleration attributed to AI assisting R&D, with heavy redactions.
Why it matters
METR spent eleven weeks inside Anthropic reviewing internal benchmarks, compute metrics and staff interviews. It concludes well-specified capabilities are likely advancing about twice the March 2024–October 2025 pace, and guesses AI now explains roughly 80% of that acceleration—while stressing AI is not yet making highly significant fully autonomous training contributions. Anthropic redacted most of the evidence METR found most informative, so outsiders must lean on the evaluator’s synthesis. Watch whether further AECI readings and less-redacted updates show acceleration compounding above 2×.
Key facts
- AECI growth ~1.9× since Oct 2025 (95% CI 1.5–2.8×)
- Majority (~80% guess) of current acceleration attributed to AI R&D help
- Not driven by faster training-compute scaling
- 11-week high-access review Aug 12–Oct 29 2026
- Anthropic redacted majority of most informative evidence
- Fully autonomous high-impact training contributions not yet observed