A lab publishes a number about itself
Anthropic published three measurements on 17 September that describe how much of its own AI research and development Claude now does. The headline figure: Claude “leads” 26% of that work as of August 2026, up from under 1% in February.
The company built what it calls an AI-Led R&D Automation Index. It catalogued the types of AI research task done at Anthropic, scored each against an automation scale running from AL0, no AI involvement, to AL5, fully autonomous, and weighted the result by how much staff time each task consumes.
“Leads” is AL4 on that scale: Claude completes most of a task end to end from a high-level prompt, with a human supervising. The rung below it, AL3, is “collaborates” — the model does large chunks under close human direction. More than 90% of Anthropic’s AI R&D now sits at AL3 or above.
No measured subset of the work has reached AL5. Anthropic states plainly that Claude is not operating fully autonomously on any of it.
Thirty thousand agents, and what watches them
The second disclosure is about oversight. Anthropic said roughly 30,000 agents were doing research and engineering work at any given moment on its main internal platform in August.
Every action those agents take passes through an online monitor before it executes. Anthropic reported that the monitor blocked 0.002% of decisions — about one in 47,000. A separate offline monitor reviews 100% of actions after the fact and flags roughly one to two transcripts per thousand. Humans directly review about 50 priority cases a week.
Those numbers are worth reading twice. A blocking rate of one in 47,000 can mean the guardrails are rarely needed, or that they catch little. The figure alone does not distinguish between the two, and Anthropic does not claim it does.
Six per cent to safety
The third measurement is compute. Anthropic said 6% of its AI R&D compute went to safety research in July 2026, and 12% of the compute specifically allocated to AI-driven R&D.
The company adds a caveat to its own figure: safety work leans on experimental design more than on raw compute, so a low share does not map cleanly onto low effort. That is a fair point and also a convenient one, and it is the kind of claim an outside evaluator would have to test.
Why now
The timing is not incidental. Five days earlier Anthropic’s chief executive Dario Amodei published an essay arguing that the industry must slow the rate at which it improves model capabilities, and that frontier labs should let independent evaluators inside to check their claims.
Publishing self-measured metrics is what that argument looks like in practice. It also sets a bar competitors have not met: no other frontier lab publishes a number for how much of its own research its models run.
What to watch
Everything here is Anthropic measuring Anthropic. The index is the company’s own construction, the automation levels are its own definitions, and nothing has been audited.
Anthropic says it intends to have independent third-party evaluators verify these metrics. The test of this disclosure is whether that happens, who the evaluators are, and whether the 26% figure survives someone else’s arithmetic.
