Anthropic published a set of internal measurements describing the pace of artificial intelligence development at its facilities, including data on how much of its AI research and development work is being performed by its Claude AI system.

According to the company's statement, as of August 2026, Claude "leads" 26% of Anthropic's AI R&D work, up from under 1% in February 2026. The company defines "leads" as the AI completing most of a task end-to-end from a high-level prompt while a human supervises. More than 90% of AI R&D work is performed at or above the level the company classifies as "AI collaborates." The company states Claude is not operating fully autonomously for any measured subset of AI R&D work.

Anthropic said it operates approximately 30,000 AI agents simultaneously on its most-used internal research and engineering platform. The company reported that 100% of those agents' actions pass through an online monitor before execution. Over August 2026, roughly 0.002% of over one billion agent decisions were blocked by the monitoring system.

On compute allocation, the company examined one week of usage between July 13 and July 20, finding that approximately 6% of compute directed toward AI R&D was allocated to safety work. About 12% of compute directed toward AI-driven AI R&D was allocated to safety.

Anthropic said it plans to embed independent third-party evaluators from multiple organizations to verify safety practices, report incidents, and monitor key metrics. The company said it intends to publish these measurements on a continuing basis and called on other frontier AI developers to disclose comparable data using a shared public methodology.

The company acknowledged several limitations in its methodology, including the difficulty of distinguishing safety research from capabilities research, reliance on best-effort workload labels, and the fact that the compute measurement covers only a single week.