Anthropic has disclosed that its own model, Claude, now “leads” 26% of the company’s AI research and development work — meaning the chatbot completes most of a task from a high-level prompt while a human supervises. The figure was below 1% in March.
The numbers come from a Thursday blog post by the Anthropic Institute, the lab’s first release of what it says will be recurring measurements of how much AI development work is being done by AI itself. The company frames the exercise as an early-warning system for recursive self-improvement: the point at which a model can fully autonomously design and build its own successor.
“We are not there yet, and recursive self-improvement is not inevitable,” Anthropic states in the underlying report. “But it could come sooner than most institutions are prepared for.”
What the metrics actually measure
The 26% figure sits on a ladder of autonomy the company has defined, from AI assisting humans to AI directing work. Alongside it, Anthropic reported that “the share of work at or above ‘AI collaborates’ is above 90%”, and that Claude is “not operating fully autonomously for any measured subset of AI R&D work” — the distinction between a model that runs a project and one that sets its own goals.
The report also disclosed the size of the internal safety net. Automated monitors reviewed more than 1 billion AI decisions in August and blocked 0.002% of them — roughly one in 47,000. About 50 high-priority cases are escalated to human review each week. Earlier internal data released in June showed that as of May 2026, more than 80% of the code merged into Anthropic’s production codebase was authored by Claude, up from low single digits before Claude Code’s research preview in February 2025.
Anthropic is explicitly inviting comparison. It says other frontier labs could publish the same figures and allow third parties to verify them. “We are reporting these measurements because they give the public, third parties, and governments better visibility into the pace of AI development inside frontier labs,” the company said.
Context: a fortnight of safety escalation
The disclosure lands in the middle of an unusually charged stretch for AI safety. Last week, Anthropic researcher Jacob Coxon resigned and accused leading AI companies of “gambling with our lives” by pressing ahead with increasingly capable models. Anthropic CEO Dario Amodei subsequently called for frontier AI companies to coordinate on slowing development — a proposal President Trump quickly rejected, and one that the site has tracked through the UK’s rejection of an AI kill switch and Congress’s Kill Switch Act push.
The metrics release is the concrete half of that argument. Anthropic’s position is that a credible slowdown would require shared, verifiable measurements — the same logic as arms control, where you cannot pause what you cannot count. On Wednesday, Geoffrey Hinton told a closed-door Capitol Hill briefing that Congress has “maybe a year” to regulate AI, citing the same recursion dynamic in front of lawmakers.
The open question
What the numbers cannot yet settle is what “leading” R&D predicts. Anthropic argues that models accelerating their own development “could make it more challenging for humans to understand or control these systems”, and that sharing metrics now lets society decide how to respond. Critics have long noted that a lab publicising its autonomy statistics while competing at the frontier is, at minimum, an unusual combination of transparency and salesmanship — and a unilateral slowdown would only change who is in front, which is why Anthropic ties the disclosure to a call for coordinated, verifiable limits.
For now, the release does something more modest but rare: it puts a number on how much of a frontier lab’s research is already being run by the product of that research. Yesterday that number was below 1%. In March it was barely anything at all. The trajectory, as much as the figure, is the story.
Sources: Business Insider; Anthropic; The Washington Post
