Nonprofit evaluator METR revealed that Anthropic may have crossed its own internal AI acceleration warning threshold back in July, with the pace of model improvement doubling over the past year. Following an eleven-week audit of Anthropic's internal research data, training metrics, and employee surveys, METR concluded that about 80 percent of this accelerated progress stems from AI being used directly in the company's own research and development.
Whether artificial intelligence can speed up the creation of better artificial intelligence serves as a critical early warning signal in the broader AI safety debate. Anthropic's Responsible Scaling Policy explicitly flags a doubling in the pace of progress due to automated research as a risk indicator that could trigger a warning level, raising concerns over potential feedback loops and rapid capability gains. Although Anthropic redacted some of the most crucial evidence from the review, METR's analysis showed that model test scores climbed roughly 1.9 times faster since October 2025, matching a trajectory where two years of progress were compressed into a single year.
While the current acceleration is primarily driven by AI acting as a research assistant rather than conducting fully autonomous development, METR found no bottlenecks likely to slow the trend over the coming year. The findings highlight the growing reality of recursive self-improvement loops within leading AI labs and underscore the urgent need for robust safety oversight as development cycles accelerate.