Claude now 'leads' a quarter of Anthropic's own R&D

Anthropic wants the public to be able to see the pace of AI development from outside the labs, and on 17 September it set out the three measurements it thinks would do the job: how much of a lab's AI research and development is done by AI itself, how well the actions of AI agents are overseen, and how compute is allocated. The proposal arrives with a snapshot taken from inside Anthropic, and the snapshot is the interesting part.

Claude leads 26% of the work

The centrepiece is the Anthropic R&D Automation Index, built by cataloguing every kind of research and development task inside the company, rating how automated each one currently is, and aggregating the results. The rating borrows Epoch AI's Automation Level scale, which runs from AL0, where no AI is involved, to AL5, where AI operates fully autonomously with nobody in the loop. AL3 is 'AI collaborates': large chunks of work done under close human direction. AL4 is 'AI leads': most of the task finished end-to-end from a high-level prompt, with a human supervising.

As of August 2026, Anthropic says Claude leads 26% of its AI research and development work and that more than 90% of the total sits at 'AI collaborates' or above. It also states plainly that Claude is not operating fully autonomously for any measured subset. The caveat sits in the same section: Anthropic is using its own models to evaluate its own work, which it acknowledges could mean the judge makes the same kinds of errors as the model it is checking. Its answer is third-party verification, and it says it plans to embed independent evaluators with access to internal systems, processes and data comparable to what its own risk teams get.

Thirty thousand agents, one monitor

The oversight measurements are the more concrete half of the post. On its most-used internal platform, Anthropic says roughly 30,000 agents were doing research and engineering work at any one time in August. Every one of their actions passes through an online monitor before it is executed, usually within seconds, and every action is ingested by a monitor afterwards; blocked actions get human review within a week. The automated monitor flags roughly 100,000 transcripts a week for classifier triage, and about 50 of the highest-priority flags each week are escalated to a person. Across more than a billion decisions in August, 0.002% were blocked, which Anthropic renders as roughly 1 in 47,000, and between one and two transcripts in every thousand are flagged for a second look.

Six per cent of R&D compute, and the number that matters

For compute, Anthropic examined a single week in July and sorted every workload into categories to ask how much of the compute going to AI research and development was safety work. The answer was about 6%, rising to about 12% for compute going to AI-driven AI research and development. Anthropic calls those deliberately conservative: if a token advanced capability as much as it advanced safety, it was not counted, and the figures exclude the classifier compute that runs alongside deployed models. Safety research is people designing experiments, which is slow and not especially compute-hungry, so compute is an imperfect proxy for emphasis — which is exactly why the useful part of the metric is that it can be compared across labs and across years.

Our opinion

Publishing the dial is not the same as turning it, and Anthropic is careful to say so. Even so, an index that puts its own model at 'AI leads' on a quarter of its research work is a genuinely uncomfortable fact to volunteer in a field where every lab talks about automation in the abstract and almost none of them publish a number. The 26% figure will be quoted for years. The methodology sitting behind it is what will decide whether quoting it means anything.

The awkward part is the same here as everywhere in AI safety reporting: the models doing the judging and the models being judged come from the same lab, and Anthropic admits that in the text rather than hiding it in an appendix. The right response is not to dismiss the number but to treat it as a claim awaiting a second opinion. If independent evaluators with real access turn up inside Anthropic and the figure survives contact with them, the industry has a template worth copying. If they do not arrive, this becomes a press release with a decimal point.