Anthropic Says Claude Now Leads a Quarter of Its Own R&D

Anthropic said on Thursday, September 17, 2026, that its Claude model now leads 26 percent of the company’s research and development work, The Washington Post reported. According to the Post, that share rose from zero in a matter of months.

The figure is Anthropic’s own description of its internal work, and the company is the only source for it. Still, it is an unusually concrete number from a frontier AI lab about how much of the job of building AI is now being done by AI.

Why the number matters

The Post framed the disclosure as a sign of progress toward AI systems that help build their successors, with the headline describing Claude as taking over the work of building its own successor. For years, the idea of AI systems contributing to their own development has been discussed mostly as a theoretical milestone. A lab putting a percentage on it moves the discussion from theory to measurement, even if the method behind the measurement has not been published in the source.

The speed matters as much as the level. A jump from nothing to roughly a quarter of R&D in months, if it holds, suggests the share could keep changing quickly. That is the kind of trend policymakers and safety researchers have said they want to track.

What “leads” may and may not mean

The source summary does not spell out how Anthropic defines a project that Claude “leads,” how the 26 percent is counted, or how much human review remains in the loop. Those details matter. Readers should treat the figure as Anthropic’s characterization rather than an independently verified statistic, and look for the company’s methodology before drawing firm conclusions.

The American policy context

The disclosure landed during a crowded stretch for US AI policy. A day earlier, a federal kill-switch bill stalled in the Senate (read our report), with Sen. John Kennedy citing concerns about AI models that improve themselves. A day later, California Gov. Gavin Newsom signed an AI safety executive order (read our report). A frontier lab saying its model now leads a quarter of its own R&D gives both debates a concrete data point.

Anthropic made another research claim a few days later, when it said Claude agents had found a previously uncharacterized enzyme system; see Anthropic’s new biology lab.

What to watch

Key questions include whether Anthropic publishes how it measures the 26 percent, whether other labs disclose comparable figures, and whether lawmakers start asking AI companies to report this kind of metric as part of safety oversight.