METR Says Anthropic Model Development Speed Has Doubled as Internal Model 4 Emerges
A METR assessment has brought Anthropic’s previously undisclosed Model 4 into public view and described a sharp acceleration in the company’s model-development pace. The report says Anthropic models improved on clearly defined tasks at twice the rate recorded across the March 2024 to October 2025 period. Model 4, trained in September and October 2025, scored 178.6 on Anthropic’s internal AECI measure, ahead of the public Opus 5.5 and Mythos 5.1 models. It cleared a 50% accuracy threshold on all 15 listed benchmarks, while removing three problematic benchmarks would raise its score to 179.7. METR estimated that about 80% of the roughly twofold research acceleration came from AI assistance rather than faster training-compute growth. Anthropic employees reported productivity gains rising from 1.3–1.4 times in spring 2025 to about 12 times while using Model 4 in October 2026. The report also said Anthropic’s public version omitted much of a roughly 100-page submission, including a redacted forecast table. METR researcher Ryan Greenblatt described the findings as a major signal for ASI within one to two years, while the report connected the discussion to the AI 2027 scenario.








