The singularity rests on one loop: machines that help build better machines, which help build better ones still. For years that loop was an argument. This autumn it became a measurement.
The number
- 26%of Anthropic's AI R&D work led by Claude, August 2026
- Under 1%the same share in February 2026
- Over 90%of the work at “AI collaborates” or above
Anthropic's new R&D Automation Index catalogues every kind of AI research and engineering work done at the company and rates how automated each task is, on a six-step scale developed by Epoch AI: no AI involvement, minimal involvement, AI assists, collaborates, leads, and fully autonomous. As of August 2026, Claude leads 26% of that work, and more than 90% of it is at “collaborates” or above.
The method is ordinary in the best sense. Every week in July, Anthropic sampled a fifth of the staff in each R&D department and listed what they actually worked on: about 15,000 tasks, sorted into a fixed tree of 542 categories. An independent model rated each task; its ratings fell within one level of a human rater's 97% of the time.
The scale of the work has changed too. At any one time, about 30,000 agents do research and engineering on Anthropic's most-used internal platform.
“AI systems are getting more powerful, and they're increasingly being used to build the next version of themselves.”
Not one lab, and not only labs
OpenAI says it has reached the goal it set last autumn: an automated research intern by September 2026. And in academia, a paper submitted on 22 September describes AIDE², a research agent that proposes changes to its own code, tests the modified versions of itself and keeps what works. In an autonomous eight-day run it found seven successive improvements, and the best agent it produced matched or beat a strong human-engineered research agent on all four held-out benchmarks.
The loop is no longer a thought experiment. It has a number and a method.
What the number does not say
No measured part of Anthropic's AI R&D is fully autonomous yet; people still direct the work. Anthropic is also candid that the index rates tasks against a July 2026 basket, so it shows existing work being automated rather than new kinds of work appearing, and that a model judging its own lab's models may share their blind spots.
That candour is the point. A lab that publishes how it measures, and what the measure misses, gives everyone a way to check the next number.
What to watch
Anthropic plans to rebuild the basket of tasks periodically and to bring in independent evaluators from other organisations. The step to follow is the one from “leads” to “autonomous”. When it moves, we will report it here, with the date.