Claude now leads 26% of Anthropic’s own AI R&D

Anthropic says Claude now leads 26% of the research and development work used to build its future AI systems, up from less than 1% in February. The company’s August snapshot also puts more than 90% of AI R&D at or above the level where Claude completes large pieces of work under close human direction.

That does not mean Claude is autonomously designing and releasing its successor. Anthropic says none of the measured work reached its highest automation level, where AI operates without a human in the loop. But its own numbers show that building frontier models is rapidly becoming a job performed by the models themselves.

From assistant to research lead

Anthropic created an R&D Automation Index by mapping the kinds of work involved in model development and rating each task on a six-level scale. At the level it calls "leads," Claude can take a high-level prompt and complete most of a task end to end while a person supervises. The company says 26% of its work now meets that description.

The index was assembled from a random sample covering 20% of relevant staff in each week of July. Claude agents turned the records into roughly 15,000 granular tasks and a frozen tree with 542 nodes, then another Claude judge assigned automation levels. Human reviewers and the model gave the exact same rating 59% of the time, while two humans agreed exactly 35% of the time. Model and human judgments were within one level 97% of the time.

The method is unusually detailed, but it remains a company measurement in which Anthropic’s models help assess Anthropic’s automation. Anthropic says outside evaluators will be embedded with access comparable to internal risk teams.

Thirty thousand agents need an oversight factory

Anthropic says about 30,000 agents were doing research and engineering work at any one time on its most-used internal platform in August. Every action on that platform passes through an online monitor before execution and is also ingested by an offline monitor afterward.

Across more than one billion agent decisions in August, the online system blocked 0.002%, or about one in 47,000. Offline monitoring flags roughly 100,000 transcripts a week for additional classifier review, with about 50 high-priority cases reaching human reviewers. The scale turns rare misbehavior into an operational problem: a tiny failure rate can still create many events when agents act billions of times.

Safety received a measurable slice of compute

For one week in July, Anthropic classified about 6% of AI R&D compute as safety work. The share rose to about 12% when measured only against compute used for AI-driven AI R&D. The company calls these estimates conservative and warns that compute is an imperfect proxy because safety research can require more human design than machine time.

Anthropic wants other frontier labs to publish comparable figures for AI-led research, agent oversight and compute allocation. The immediate oddity is not full recursive self-improvement. It is that one lab can already count tens of thousands of research agents, measure how often they are stopped, and say a quarter of the work building their successors is led by the models themselves.