Anthropic starts measuring how much AI runs its own research

Anthropic proposes three metrics that would let outsiders track how fast development inside a frontier lab is moving, and it published its own current readings to show what they look like. The first measures how much of the lab's AI research is done by AI. Anthropic says Claude now leads 26% of its AI research and development tasks, up from under 1% in February 2026, and works at a "collaborates" level or higher on more than 90% of them.

The second metric covers oversight of the roughly 30,000 agents running inside the company. Every agent action is monitored, either checked before it runs or reviewed after. Automated review happens in seconds, human review within a week. Online monitors block about 0.002% of decisions and flag roughly 0.1% for a closer look. The third metric is compute allocation: in their snapshot, about 6% of AI R&D compute went to safety work, and 12% of the compute driven by AI itself was directed at safety.

The point of publishing these, Anthropic argues, is to connect inputs like compute to outputs like capability, so the public and other labs can see the production process instead of guessing at it from model launches. It also frames the numbers as a starting point for coordinating on pace across labs. You can read the proposal for the full definitions.

Why it matters

If you follow AI policy or safety, these are among the first concrete internal numbers a lab has put on the table, which gives the debate something specific to argue about instead of impressions. If you work at a lab, expect pressure to report comparable figures, because a single published baseline tends to become the standard everyone is measured against.