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

Anthropic has started publishing a self-measurement of how much of its own AI research is being done by AI. In a 17 September 2026 post from the Anthropic Institute, the company said Claude now "leads" 26% of Anthropic's AI research and development work as of August 2026 — up from under 1% in February. It is the first release of what Anthropic calls a prototype R&D Automation Index.

What the numbers mean

The index scores tasks on an automation scale developed by the nonprofit Epoch AI, running from AL0 (no AI involvement) to AL5 (AI works fully autonomously, no human in the loop). At AL3, Claude "collaborates" — doing large chunks of work under close human direction. At AL4, "leads," it can complete most of a task end-to-end from a high-level prompt while a person supervises the result. Anthropic says more than 90% of its AI R&D work now sits at AL3 or above, but that no measured work has reached full autonomy at AL5.

How it was measured, and its limits

For one week in July 2026, a Claude research agent reviewed a random 20% sample of staff across every department in the model R&D loop, cataloguing roughly 15,000 tasks and sorting them into a fixed taxonomy of 542 categories. Anthropic itself flags the caveat that matters most for outside readers: model and human graders agreed on the exact automation level only 59% of the time, versus 35% agreement between two human graders, though the two landed within one level of each other 97% of the time. The company also reports around 30,000 concurrent agents doing internal research and engineering work, with automated monitors blocking roughly 1 in 47,000 agent decisions.

Why the framing matters

This is a vendor measuring itself, using a model it built to grade the work, and publishing the result days after its own CEO argued publicly for slowing AI development industry-wide. Anthropic frames the index as a template any frontier lab could adopt, not an audited industry benchmark. Buyers evaluating AI-automation marketing from any vendor should ask for the same things Anthropic disclosed here — method, sample, and grading caveats — rather than accepting a headline percentage on its own.

Sources: Anthropic Institute, Unite.AI.