The announcement
Anthropic published three proposed measurements for the public to follow: how much of its AI research Claude helps do, how its internal agents are monitored, and how research computing is divided between different kinds of work. The company says that, as of August, Claude led 26% of its measured AI research tasks under human supervision. It also says Claude was involved in more than 90% of that work at a collaboration level or higher. None of the measured work was fully autonomous.
Why it matters
A model helping build its successor sounds dramatic, but “involved” and “fully in charge” mean very different things. The useful question is how much work a person still directs and checks, and whether the company can show that its oversight is keeping up. These are Anthropic’s own measurements from a snapshot of its internal work. The company proposes outside evaluation; the numbers are not yet a shared, independently verified industry standard.