Claude Now Leads 26% of Anthropic’s AI R&D Work

Artificial intelligence is no longer being used only to help people write emails, generate images, or analyse spreadsheets.
At the frontier of AI research, AI is increasingly participating in the process of building the next generation of AI itself.
Anthropic has released new measurements showing that, as of August 2026, Claude was leading 26% of the company’s measured AI research and development work. More than 90% of the measured work involved AI at or above what Anthropic describes as the “collaborates” level.
That 26% figure needs an important qualification.
Anthropic is not saying Claude has become an autonomous scientist operating without humans.
Its measurement framework ranges from AL0, where there is no AI involvement, through to AL5, where an AI operates fully autonomously with no human in the loop.
Claude has not reached AL5 in any measured subset of Anthropic’s R&D work.
At AL4 — the level Anthropic calls “leads” — an AI can complete most of a task from a high-level prompt, but a human still supervises the process.
That distinction may become one of the most important distinctions in AI development.
The conversation is changing from:
“Can AI help a researcher?”
to:
“How much of the research workflow can AI lead while humans retain meaningful oversight?”
This creates a fascinating feedback loop.
Humans build increasingly capable AI.
That AI helps humans conduct AI research.
The improved research contributes to the development of stronger AI.
Those stronger systems may then contribute even more heavily to the next research cycle.
That is not necessarily uncontrolled recursive self-improvement. Humans still provide objectives, infrastructure, access, evaluation, and deployment authority.
But it is a form of accelerating research feedback.
And that makes the human layer even more important.
When AI can perform more of the execution, humans increasingly become responsible for something harder: judgment.
What should be researched?
Which result should be trusted?
Which experiment should proceed?
What requires independent verification?
Where must humans retain authority regardless of how capable the system becomes?
Anthropic is also proposing that frontier AI developers consider publishing comparable measurements of AI involvement in their R&D.
That could eventually give researchers, governments, and the public a more meaningful measurement than simply asking whether a company “uses AI.”
We may eventually ask:
How much of your AI is being built with the assistance of your previous AI?
That is a very different question.
In 2026, we are beginning to get numerical answers.