Claude now leads 26% of the work building its own successor

In February, Claude led none of Anthropic's model research. By August it led 26%. Anthropic said this week that Claude is now helping build the next, smarter version of itself.

"Leading" has a narrow meaning here: Claude can finish most of a task end-to-end from a high-level prompt while humans supervise. Another measure is broader. About 90% of Anthropic's R&D now happens "in collaboration" with Claude, with the model doing large chunks of work under close human direction. The company says about 30,000 agents were doing R&D work as of August, and it recently agreed to outside safety evaluators.

Anthropic calls this progress toward recursive self-improvement, a model building its own successor on its own, but says it hasn't reached that point. It also warns that models speeding up their own development could become harder for humans to understand or control. It wants the industry to measure this publicly and report on it.

The timing matters. A day later, California Gov. Gavin Newsom signed an executive order giving experts 60 days to recommend stronger AI safety rules. One idea on the table is an emergency shutoff for frontier models.

Why it matters: going from 0% to 26% in six months is a steep curve, and it's the kind of number regulators will now expect every lab to report.