METR Audit Finds Anthropic's Model Development Speed Doubled, Hinting at AI Self-Acceleration

A report by non-profit METR, based on an 11-week internal investigation of Anthropic, finds that the company's model capabilities are improving about twice as fast as a year ago, with roughly 80% of the acceleration attributable to AI being used in Anthropic's own R&D rather than increased training compute. This pace may have already crossed Anthropic's self-set risk warning threshold, but the company redacted key evidence, limiting independent verification.

The report reveals three unreleased internal models, Model 2/3/4, with the strongest Model 4 scoring 178.6 on Anthropic's AECI index—about 3.9 points above the public Opus 5.5, equivalent to roughly three months of normal progress. METR estimates the rate of capability improvement since October 2025 is about 1.9 times faster (95% CI 1.5–2.8x), with no clear bottlenecks in algorithms, compute, or data over the next 12 months. Employees report Model 4 delivers about 12x productivity gains, and there are signs of AI independently handling key training work.

The findings have raised alarms about the risks of uncontrolled R&D automation, with METR's director urging society to prepare for rapid change. Anthropic co-founder Sam McCandlish acknowledged that AI is accelerating its own R&D, but the public report omits many details for commercial sensitivity and to avoid providing an acceleration roadmap. If the trend continues, self-accelerating AI R&D could outpace social and political adaptation, and existing evaluation methods face saturation risks.