Anthropic proposed new metrics to give the public visibility into frontier AI development. It shared a snapshot of these metrics from inside its own labs. The metrics cover AI-led R&D, oversight of AI agents, and compute allocation. Anthropic measured how much of its AI R&D is performed by AI. It found AI leads 26% of its AI R&D work. Over 90% of its AI R&D work is at or above the level where AI collaborates. Anthropic also measured oversight of its AI agents. It has 30,000 agents active at once on its most-used internal platform. 0.002% of agent actions were blocked by its monitor. Anthropic measured how it allocates compute for AI R&D. 6% of AI R&D compute went to safety work. 12% of AI-driven AI R&D compute went to safety work. Anthropic plans to embed third-party evaluators to verify these metrics. It will give them access to internal processes, systems, and data.

• 26% of AI R&D work led by AI
• 90%+ of AI R&D work at or above AI collaborates
• 30,000 agents active at once
• 0.002% of agent actions blocked