The Pulse
Anthropic Says Claude Leads 26% of AI Research, None Autonomously
Anthropic’s prototype index puts Claude at the “leads” level for 26% of its AI research and development work, while reporting no fully autonomous AI R&D. The company’s account also describes gains in coding and research tasks, alongside are
AI.info Team · Written with AI · fact-checked ·
“Claude is not operating fully autonomously for any measured subset of AI R&D work.”
Anthropic, in its measurement of AI R&D automation
Anthropic’s research and development automation index puts Claude at the “leads” level for 26% of the company’s AI research and development work. The company’s measurement page reports that figure as of August 2026 and says more than 90% of the work is at “AI collaborates” or above. It also says Claude is not operating fully autonomously for any measured subset of AI R&D.
Anthropic describes recursive self-improvement as an AI system fully autonomously designing and developing its own successor. The index uses an automation scale developed by Epoch AI, from AL0, with no AI involvement, to AL5, where AI operates fully autonomously without a human in the loop. At AL3, AI can complete large chunks of work under close human direction. At AL4, AI can complete most of a task end to end from a high-level prompt while a human supervises.
The 26% figure is an estimate of work at the “leads” level, not a count of AI systems building new models without people. Anthropic says its prototype index catalogs the kinds of AI R&D work done at the company, rates how automated each task is, and aggregates those ratings. The company presents the index as a way to measure how AI participates in building future AI systems.
AI takes on more of the development work
In its separate post “When AI builds itself,” Anthropic describes a growing role for AI in its development process. It says more than 80% of the code it merged into its codebase was authored by Claude as of May 2026. In the second quarter of 2026, the typical engineer was merging eight times as much code per day as in 2024. Anthropic cautions that lines of code are an imperfect measure of productivity because they track quantity rather than quality.
The post also reports improvements in tasks that require Claude to work without step-by-step instructions. On open-ended coding problems, Claude’s success rate reached 76% in May 2026, up 50 percentage points in six months. Anthropic defines session success as the agent clearly completing the user’s task without corrections, based on a judgment by Claude. The company notes that changing workloads can cause short-term fluctuations in the rates.
Anthropic describes a research experiment in which Claude-powered agents worked on an open problem in AI safety: whether a weaker model could reliably supervise a stronger one. The agents proposed hypotheses, tested them, shared results with parallel agents and iterated. They recovered 97% of the gap between the task’s performance floor and ceiling over 800 cumulative hours. Two human researchers recovered about 23% of the gap over roughly a week. Anthropic says the result did not transfer cleanly to production-scale models, and that humans chose the problem and created the scoring rubric.
Human judgment remains part of the process
Anthropic’s account distinguishes performing research tasks from deciding which problems to pursue. In its assessment, Claude can execute well-specified experiments and is improving at suggesting next steps, but gaps remain in judgment about goals. In one analysis of selected research-session moments, the company found that its best model beat the human’s choice of next step 64% of the time in April 2026, compared with 51% for its best model in November 2025. Anthropic cautions that the moments were deliberately selected because the human’s decision had room for improvement, so the comparison is not like-for-like.
The two measures describe different parts of the picture: the index places a portion of Anthropic’s R&D work at a level where Claude leads under human supervision, while the recursive self-improvement post describes gains in coding and research tasks. Neither report says Claude has autonomously built its own successor. Anthropic identifies choosing which problems matter, judging results and deciding when an approach has failed as areas where human judgment still matters.