Claude takes the lead on 26% of Anthropic's R&D

Anthropic said Claude now leads 26% of the company’s own model research and development, up from 0% in February. “Leading” means the model completes most of a task end to end from a high-level prompt, with humans still supervising the result — distinct from “collaborating,” where a person drives and Claude assists. About 90% of all R&D at Anthropic now happens in some form of collaboration with Claude, and only a sliver (Anthropic has disclosed the number as essentially zero) of tasks run fully autonomously with no human check on the output.

What that actually looks like

The work spans concrete engineering tasks rather than abstract strategy: implementing model improvements, reviewing code, keeping documentation in sync, and running evaluations. Anthropic published the numbers as part of a transparency push tied to its proposal for measuring AI development pace — the same metric, “AI performing AI R&D,” is one of the three proposed measurements that labs would report publicly.

Why it’s a bigger deal than it sounds

A quarter of R&D being led by the model itself is the point at which “we’re the lab doing the work” stops being a strictly human claim. If a frontier lab keeps scaling this, the trajectory of models becomes partly self-running — which is exactly why Anthropic is pushing for these numbers to be measured and reported with the same rigor across labs, rather than left to occasional disclosure.

(Original source: Anthropic)