METR Estimates Anthropic’s AI Benchmark Improvement Rate at 1.9 Times Historical Pace

Anthropic’s AI benchmark improvement rate has accelerated to about 1.9 times its earlier pace since October 2025, with most of the increase tied to internal research use, METR estimated.

METR compared the improvement rate since October 2025 with the pace from March 2024 through October 2025. The comparison used a benchmark METR calls AECI.

The review covered 11 weeks, from Aug. 12 to Oct. 29. METR examined confidential data, interviewed Anthropic staff and reviewed proprietary performance metrics.

About 80% of the acceleration came from Anthropic’s internal use of its models for research and development. Most of the gains came from models assisting human researchers rather than running processes fully autonomously. Faster growth in training compute did not drive the increase.

The assessment follows Anthropic’s Oct. 15 announcement that six METR employees were gaining access to certain internal systems, equipment and office space through its safety review arrangement.

The findings bear on concerns about recursive AI self-improvement, in which AI systems help develop later systems. METR’s estimate is based on one benchmark and an 11-week evaluation period.