Anthropic says Claude now leads 26% of its own AI research

Anthropic has put a number on how much of its own AI is being built by AI. As of August 2026 the figure is 26%, and six months earlier it was under 1%.

What the index measures

The number comes from the first results of Anthropic’s R&D Automation Index, published September 17, 2026. It says Claude now leads 26% of the company’s AI research and development work.

The definition carries the weight. Leading means the model completes most of a task end to end from a high-level prompt while a human supervises, which is a specific rung rather than a general impression of usefulness. Anthropic states that Claude is not operating fully autonomously on any measured subset of the work. The rung above, where no human is in the loop, has not been reached.

A second figure gives the wider shape: more than 90% of the measured work now sits at the collaborates level or above. The index scores each task against an automation rating scale developed by Epoch AI, and Anthropic says it plans to re-version the published numbers periodically as it rebuilds the underlying basket of tasks.

Why the disclosure is the story

The trend line matters more than the level. Going from under 1% to 26% in roughly six months is the kind of slope that makes any single reading obsolete quickly, and it describes the recursive case specifically: models contributing to the research that produces the next models.

The measurement arrives days after the same company called on the industry to pace itself, which puts it in the position of documenting the acceleration it says needs restraint. Anthropic names two obstacles to comparing this kind of reporting across labs: the lack of a common methodology, and the fact that it is using its own models to evaluate its own systems.

That is the limitation worth holding onto. The index is self-measured, self-defined and self-published, and its credibility rests on Anthropic continuing to publish it when the number becomes less comfortable.