Anthropic’s AECI benchmark scores have improved faster since October 2025, with Model Evaluation and Threat Research (METR), an AI research nonprofit, estimating a 1.9-fold acceleration while cautioning that its cause remains uncertain.
METR measured the change on Anthropic’s AECI, its version of the Epoch Capabilities Index, which tracks performance on verifiable tasks. The 1.9 figure came with a 95% confidence interval of 1.5 to 2.8, while methodological variations produced estimates from 1.1 to 2.9. Put in historical terms, the estimated pace would compress roughly two years of earlier benchmark gains into one year.
Separately, METR judged that, as of October 2026, Anthropic’s models were improving on well-specified tasks at about twice the pace recorded from March 2024 to October 2025. The group said a rate closer to 1.3 times or three times the earlier pace would not surprise it.
METR said it lacked decisive evidence about AI’s role, and Anthropic staff did not agree on how much AI contributed. The nonprofit estimated that AI might account for about 80% of the current acceleration, with substantial uncertainty. It also judged that factors unrelated to AI probably contributed to late-2025 gains, while faster growth in training-compute scale did not explain the acceleration.
Informal internal Slack polls included staff estimates as high as 12 times in October 2026. METR gave the responses limited weight because they reflected little deliberation. It said AI systems had yet to carry out major, high-impact model-training tasks independently.
The review ran from Aug. 12 to Oct. 29, 2026, and drew on internal benchmark and training data, interviews and other company information. METR’s team totaled 2.5 full-time equivalents. Anthropic could decline requests and redact nonpublic material; METR said the company redacted most of the evidence it considered most informative.
METR cautioned that AECI results may not reflect progress on broader, less-defined work. It also said benchmark overfitting or Anthropic’s comparatively slow historical growth could have pushed its estimate higher.
