METR estimates Anthropic’s AI progress accelerated 1.9-fold

METR estimated that Anthropic’s AI models have improved on benchmark tasks about 1.9 times faster since October 2025 than during the period from March 2024 to October 2025, and said AI assistance within the company’s research may account for much of the acceleration.

The nonprofit’s assessment, based on an 11-week review of Anthropic data and research processes from August 12 to October 29, put the benchmark acceleration’s 95% confidence interval at 1.5 to 2.8 times. METR said its broader estimate of progress on well-specified tasks was about twice the earlier pace. The findings were based on internal metrics, interviews and performance data, including results from unreleased models.

METR estimated that AI use in Anthropic’s research accounted for about 80% of the acceleration, but stressed that the estimate was uncertain. Anthropic researchers did not agree on AI’s contribution; other possible factors included new hires and changes unrelated to AI. METR also cautioned that benchmark performance on clearly defined tasks may not reflect progress on less structured research work.

The assessment found that faster growth in training compute did not explain the higher pace. METR said...