Anthropic published an R&D Automation Index showing that Claude now “leads” 26 percent of the work used to build its successor models, up from under 1 percent in February. The company said it will publish the metric regularly so outsiders can see how quickly AI is taking over AI research.
That number is the closest the industry has come to a public gauge of recursive improvement inside a frontier lab. It does not mean Claude is unsupervised. It does mean a quarter of the research workflow is now model-led rather than human-led.
The disclosure landed beside Amodei’s pacing argument and OpenAI’s own notes on research acceleration. If models are already doing a large share of the work that produces the next models, calls to slow the frontier are also calls to slow a feedback loop that labs themselves have switched on.
Competitors will copy the measurement even if they dispute the 26 percent figure. A shared automation index would give policymakers something more concrete than essays when they ask how fast the stack is eating itself.
Key takeaway. Automated research is no longer a thought experiment. Anthropic has put a public number on how much of its next model is being built by the current one.
Photo: Unsplash / National Cancer Institute (lab research). Sources: Anthropic disclosures via Reuters, Los Angeles Times, AIToolsRecap, Sept. 17–19, 2026.
