An audit by safety research nonprofit METR found that Anthropic’s rate of model improvement roughly doubled since October 2025, driven largely by AI systems assisting in the company’s internal research. METR estimates that approximately 80 percent of this acceleration stems from using AI as an assistant in development workflows, allowing researchers to automate setup steps and iterate faster.
According to METR, Anthropic’s model benchmark scores climbed at 29 AECI points per year compared to 15 points in the prior period, accomplishing in one year what previously took two. However, the finding raises safety questions under Anthropic’s Responsible Scaling Policy, which flags rapid pace increases from automated AI research as a risk indicator for potential feedback loops and emerging capabilities.
Why it matters
Highlights the rapid emergence of recursive AI research acceleration using automated assistants.
Triggers internal risk thresholds for Responsible Scaling Policies regarding recursive self-improvement.
Signals rising auditability standards for frontier labs using independent external evaluation organizations.
Source: the-decoder.com



