Anthropic published metrics showing Claude led 26% of internal AI research and development in August, without fully autonomous operation.

Anthropic’s disclosed metrics show that Claude led 26% of its internal AI research and development in August 2026, compared with almost zero at the beginning of the year. More than 90% of measured work reached at least the “AI collaboration” level (AI 协作), but no measured subset of AI research and development had reached fully autonomous operation.

Approximately 30,000 agents were simultaneously performing research and engineering work internally during August. Anthropic reported that online monitoring blocked approximately 0.002% of agent decisions, described in the source as roughly one in 47,000 decisions; this is a monitoring-intervention metric, not a reported incident rate.

For July 13–20, approximately 6% of Anthropic’s AI research-and-development compute was allocated to safety work. Within the distinct pool of compute used for AI-driven AI research and development, approximately 12% went to safety-related tasks during that same week.

The metric framework covers AI participation in developing subsequent models, oversight of agent behavior, and resources devoted to more capable models. Anthropic also reiterated plans to bring in independent third-party evaluators with access to internal processes, systems, and data comparable to employees’ access, to verify safety practices, report incidents, and monitor key indicators.