Anthropic says Claude leads 26% of its R&D

Anthropic says Claude leads 26% of its AI R&D work, while more than 90% reaches the “collaborates” level or higher and none of the measured work is fully autonomous. Anthropic’s measurement report defines “leads” as completing most of a task end-to-end while a human supervises; its prototype index sampled roughly 15,000 model-R&D tasks from July and warns that judge-model agreement and self-reported labels limit cross-lab comparison. Research teams should file 26% as a supervision-and-process signal, not a productivity benchmark, and compare oversight cost with the archive’s $600 agent-day analysis.