❯ Anthropic’s internal Model 4 scores above Opus 5.5 in METR assessment

The strongest internal model surpasses Opus 5.5A November 5 assessment by METR reveals three unreleased Anthropic models. Model 2, Model 3 and Model 4 completed training in July, August–September and September–October, respectively. Model 4 scores 178.6 on the AECI capability index, about four points above Opus 5.5. At the earlier pace of roughly 15 points a year, that represents about three months of progress. The index combines internal and external tests to track changes in capability.

Progress runs at about 1.9 times the earlier paceMETR estimates that progress on these well-specified tasks has accelerated to about 1.9 times its previous rate since October 2025, with a 95% confidence interval of 1.5–2.8 times. Faster growth in training-compute scale did not explain the change. AI-assisted research likely accounts for much of the acceleration, although its precise contribution remains substantially uncertain.

Less time from idea to experimentOne interviewed researcher estimated that AI assistance cut the time from an idea to a running experiment by about 80%. Other interviews described less time spent implementing code and more experiments running in parallel. METR says major contributions to the most important training stages still primarily involve AI assisting researchers.

Anthropic plans its own analysisAnthropic redacted substantial commercially sensitive evidence, limiting public scrutiny of the conclusions. Co-founder Sam McCandlish acknowledged the cost of those redactions and plans to publish the company’s own analysis of research progress and its relationship to the Responsible Scaling Policy in the coming weeks.

▮ SIGNALLess time spent coding and setting up experiments lets a research team test more ideas, extending AI’s contribution into the model-development process itself.