METR finds Anthropic's models improving about twice as fast, with AI likely the main reason
The independent evaluator also says Anthropic's August risk report understated the speed-up, and notes that key evidence was redacted.
From August 12 to October 29, METR staff carried out what they call a high-access evaluation at Anthropic. They reviewed internal pretraining and reinforcement-learning metrics, compute figures, benchmark scores for unreleased models and forecasts, and interviewed staff. The report, published Thursday, concludes that as of October Anthropic is likely improving its models on well-specified tasks at around twice the rate it did from March 2024 to October 2025. METR flags high uncertainty and says a true figure of 1.3x or 3x would not surprise it. On Anthropic's internal capabilities index, benchmark scores have grown about 1.9x faster since October 2025 (95% interval 1.5–2.8x), roughly two years of past progress in one.
METR attributes most of the speed-up to AI itself, perhaps 80% according to Reuters, and says it is confident the cause is not faster growth in training compute. It found no convincing evidence of bottlenecks in algorithms, compute or data that would slow progress over the next 12 months, and warns of a significant risk that acceleration gets even faster over the coming year. It adds that current AI cannot yet make highly significant, fully autonomous contributions to the most important parts of model training.
In a separate review of the August 2026 Risk Report, METR agrees that catastrophic risk from models available on July 15 was low. However, it calls the report's claim that the 2x early-warning threshold had not been crossed overconfident, says it is plausible the threshold had already been crossed, and says the public report understates the acceleration. METR notes that some of the findings that most shaped its conclusions were redacted by Anthropic. The report follows this week's S-1/A language saying Anthropic may 'imminently' meet its own threshold for dramatically accelerated, AI-driven progress.
- Estimated acceleration
- ~2x (range 1.3x–3x)
- Benchmark growth since Oct 2025
- 1.9x (95% CI 1.5–2.8x)
- Share attributed to AI
- Majority, perhaps ~80%
- Assessment window
- Aug 12 – Oct 29, 2026