The Maker of the Claude Chatbot Says Its AI Now Leads a Quarter of the Work of Building Its Next AI

Anthropic published its first hard numbers on how much of its own research is done by its own AI: Claude now leads 26 percent of the work of building new models, up from under 1 percent in February, with roughly 30,000 AI agents working at any one time. Humans still supervise everything, and the company says the figures are rough, but it is publishing them because the public "needs more information" about how fast this is moving.

Anthropic, the company that makes the Claude chatbot, published a report today measuring something the AI industry has talked about for years but never put numbers on: how much of the work of building the next AI is now done by the current one.

The headline figure is 26 percent. That is the share of Anthropic's own AI research and engineering tasks in which Claude "leads," which the company defines as completing most of a task from a short instruction while a person supervises. In February that number was under 1 percent. More than 90 percent of the work now happens at a level where the AI at least "collaborates," meaning it does large pieces under close human direction. And at any given moment, about 30,000 AI agents, copies of Claude given a job and the tools to do it, are working on research and engineering inside the company.

Anthropic is careful to say what the numbers do not show. "Claude is not operating fully autonomously for any measured subset of AI R&D work," the report says. A person is still in the loop on everything. The company also lists its own weak spots: Claude was used to grade how automated Claude's work was, and when humans checked, they agreed with it only 59 percent of the time. Some of the figures come from a single week in July. And the definitions of what counts as "safety" work versus everything else are, in the company's words, "one reasonable choice among many."

Two other measurements are in the report. One tracks how closely those 30,000 agents are watched: an automatic monitor blocks about one in every 47,000 agent actions, and a slower review flags one or two transcripts in every thousand for a human to read, which works out to roughly 50 escalations a week. The other tracks where computing power goes: about 6 percent of the computing used for AI research went to safety work, or 12 percent if you count only the research that AI itself was doing.

Why publish this at all? The report's own answer is blunt. "AI systems are becoming exponentially more powerful and have begun to automate more of the process of building themselves," it says, and the public "needs more information." It also notes that when there are millions or billions of agents in the economy, "even rare events can happen regularly." The company frames the numbers as a starting point that other labs could match, so that outsiders can compare them over time.

For people who worry that AI could one day improve itself faster than anyone can follow, this is the first real data point from inside a frontier lab, and it cuts both ways. A quarter of the work, up from almost nothing in seven months, is a steep curve. A person still watching every task is the reason it is not yet the thing they fear. The report measures exactly the distance between those two facts.