Claude Is Involved in 90% of Anthropic's AI Research. Humans Still Make 85% of the Final Calls.

Anthropic says Claude collaborates on more than 90% of its AI research and development work and leads 26% of it, but that no measured subset of that work is fully autonomous. In other words, Claude is deeply embedded in how the lab builds its models, and humans are still the ones signing off.

The numbers describe participation, not an independent AI researcher. Nothing in Anthropic's measurements shows Claude setting research priorities, validating its own findings, or approving work for further development.

The Decoder reports more granular figures from an analysis of 769 task logs from model development. AI agents contributed as much as 55% of method proposals on some tasks, while humans made more than 85% of final decisions. The Associated Press describes the same picture: Claude can complete many tasks from a high-level prompt, but under close human supervision.

The most interesting finding is that about one-third of the tasks would not have been attempted without AI assistance. That suggests agents don't just speed up existing work, they widen what a research team is willing to try. But the data doesn't show whether that extra work produced better research, only that more of it happened.

These are Anthropic's own internal measurements of its own workflows. They are not an industry benchmark, and they don't show the same pattern would hold at other labs.

Bottom line: Claude has become routine infrastructure for AI development at the company that built it, which is a real milestone. But the same data shows people still making the final calls. More AI involvement means more output to review, not less accountability for reviewing it.