Written by Admin Alex · Fact-Checked by M.Ali · Info Verified September 2026
We review and update this article regularly as new information becomes available.
TL;DR: Anthropic disclosed that Claude now leads roughly 26% of the company’s model research and development work, up from essentially zero in February. The company says a human still supervises every step, but it also acknowledged that AI systems accelerating their own development could make them “more challenging for humans to understand or control.”
Zero to 26% in about six months is a fast curve for any metric, and this one happens to measure how much of Anthropic’s own model development is now being led by Claude rather than by Anthropic’s human researchers. The company disclosed the figure this month, framing it as a milestone in AI-assisted research rather than a warning sign, though it’s not hard to read it as a bit of both.
Roughly 30,000 AI agents are reportedly doing research and engineering work inside Anthropic right now, with humans monitoring rather than directing most of the individual tasks. Anthropic says about 90% of its R&D happens in what it calls collaboration with Claude, meaning a human sets the direction and Claude does large chunks of the actual implementation work, often completing tasks “end-to-end from a high-level prompt” with a person checking the output rather than writing each step.
Why this matters beyond Anthropic
This is the concept AI researchers call recursive self-improvement: an AI system that gets meaningfully better at building the next version of itself, showing up here in a very early and still human-supervised form. It’s one of the scenarios that shows up most often in AI safety literature as the point where progress could stop scaling steadily and start scaling in a way that’s much harder to predict or slow down.
Anthropic isn’t hiding from that framing. The company said outright that models accelerating their own development could make it “more challenging for humans to understand or control these systems,” a notably blunt admission for a company also actively building the thing it’s describing as harder to control. It’s the kind of statement that reads differently depending on whether you think Anthropic is being responsibly transparent or building the thing anyway and hoping disclosure counts as a safeguard.
Anthropic isn’t the only lab leaning into AI-assisted AI development. Google DeepMind and OpenAI have both talked publicly about using their own models to speed up research, write code, and design experiments, though neither has put a specific percentage on how much of their R&D pipeline is now AI-led the way Anthropic just did. That specificity is arguably the most useful part of this disclosure: a real number, even an unsettling one, tells outside observers more than another round of vague statements about AI helping researchers.
The oversight Anthropic says is still there
The company points to a few specific guardrails: constant human supervision rather than fully autonomous operation, and a commitment to bring in external third-party evaluators to monitor its safety efforts independently rather than grading its own homework. Whether that oversight scales at the same pace the R&D percentage has been climbing, from zero to a quarter of all research work in six months, is the actual open question here, not whether Anthropic means well.
Bottom Line: A quarter of Anthropic’s own model research is now led by its own AI, and that number was zero seven months ago. Human supervision is still in the loop today. The pace of that curve, not today’s snapshot, is what should hold everyone’s attention here, including Anthropic’s.
