- METR staff conducted a high-access assessment inside Anthropic from August 12 to October 29, 2026, to measure whether AI is accelerating Anthropic’s own AI research.
- The report concludes Anthropic is improving model capabilities at roughly twice the rate it did between March 2024 and October 2025.
- METR attributes about 80% of that acceleration to Anthropic’s heavy internal use of AI in its own R&D process, not to larger training-compute budgets.
- Anthropic gave METR access to pretraining, reinforcement-learning, and compute data, internal benchmark scores, speedup analyses, model forecasts, and staff interviews.
What Happened
METR published an independent assessment on November 5, 2026 of how much AI is speeding up Anthropic’s own model development, after an 11-week on-site evaluation that ran from August 12 through October 29.
Why It Matters
This is METR’s second major look inside a frontier lab’s development process this year; the group’s earlier evaluation found it could barely measure the capability gains of a prior Anthropic model. The new finding gives one of the first independently reviewed estimates of how much labs’ own AI tools are now driving their research speed, rather than relying on labs’ self-reported claims.
Technical Details
Anthropic granted METR unusually broad access: metrics on pretraining and reinforcement-learning runs, training compute broken out by domain, benchmark scores for unreleased internal models, AI-driven speedup analyses, forecasts for upcoming models, and interviews with Anthropic staff. METR found the roughly 2x acceleration in capability improvement is not explained by a corresponding jump in training-compute scale.
Who’s Affected
The assessment matters most for AI-safety researchers and regulators trying to track recursive self-improvement risk, since an 80%-AI-driven acceleration in a lab’s own R&D is a concrete data point rather than a hypothetical. It also bears on competitors like OpenAI and Google DeepMind, who face pressure to demonstrate comparable internal efficiency gains.
What’s Next
METR says the assessment relied on data Anthropic chose to share and interviews it arranged, a limitation METR flags itself; independent replication at another lab would help confirm whether the roughly 2x acceleration pattern holds industry-wide.
