Anthropic Models Gained Capabilities Nearly Twice as Fast by October

Anthropic’s models were likely gaining capability on well-defined tasks at about twice their earlier pace by October 2026, with AI tools in research and development likely accounting for most of the acceleration.

In an 11-week review of Anthropic’s internal research data, unpublished model test results and employee interviews and surveys, METR estimated that capability progress on tasks with clear scoring criteria was about twice the pace from March 2024 to October 2025. The estimate ranged from 1.3 to three times that earlier rate. On a separate benchmark, Anthropic’s version of the Epoch Capabilities Index (AECI), score growth was about 1.9 times its previous pace, with a 95% confidence interval of 1.5 to 2.8 times.

The benchmark measures performance on tasks with defined scoring rules. METR cautioned that test selection and calculation methods affect the results, and that test scores do not necessarily translate directly to real-world capabilities.

Training compute did not grow faster alongside the model gains. Anthropic’s internal research data and employee feedback pointed to AI tools helping researchers improve algorithms and design experiments, which METR assessed as likely accounting for most of the acceleration. Employee estimates of AI’s contribution varied and carried significant uncertainty, while the most important research work was not fully automated.

The safety question concerns Anthropic’s twofold R&D acceleration warning threshold. The company’s August risk assessment concluded that AI had not doubled its R&D pace. Using values from Anthropic’s private August report, METR calculated one measure at 1.97 times; that figure does not appear in the report. METR said a threshold crossing by July 15, 2026, was plausible, although most METR researchers and relevant Anthropic staff leaned toward the threshold not having been crossed.

METR found no evidence of significant constraints from limits on algorithmic progress, training compute, experiment compute or data that would prevent further progress over the next 12 months. It said it was not confident in that assessment and might have missed bottlenecks. METR said Anthropic redacted much of the evidence it considered most valuable from METR’s public assessment, limiting readers’ ability to assess its conclusions.

Anthropic has also developed a prototype index for tracking AI’s role in its research and development, including whether people can oversee that work as automation increases.