Claude went from 1% to 26% of Anthropic's model R&D in five months — measured by Anthropic

Disclosure: this newsroom's assistant is built by Anthropic, the subject of this post.

Anthropic published three measurements on 17 September intended to let people outside a frontier lab see how fast work inside one is moving.

Automation. As of August 2026, Claude was at the "leads" level for 26% of Anthropic's model R&D, up from 1% in March, and at collaborator level or above for more than 90%. The index required building a map of all AI R&D tasks at the company, a scale for how automated each is, and a weighting — constructed from Anthropic's own work records, including Slack and internal documentation.

Agent oversight. About 30,000 agents do research and engineering work at any one time on the company's most-used internal platform. Every action passes an online monitor before it runs; of more than a billion decisions in August, 0.002% — roughly 1 in 47,000 — were blocked.

Safety compute. In the week of 13–20 July, about 6% of AI R&D compute went to safety work, rising to about 12% of the portion that AI was itself carrying out.

The tracker records the disclosure.

Why it matters

Analysis — interpretation, not additional fact.

Take the numbers first, because they deserve it. 1% to 26% in five months is the most concrete public figure anyone has put on recursive self-improvement, and it is exactly the thing Amodei cited as a trigger nine days ago without attaching a number to it. Now there is a number. Whatever else is true of this disclosure, a lab publishing a curve of how much of its own research its model is doing is more useful than a lab not publishing one, and no other lab publishes one.

Then take the problem, which is not a small one. Every figure here is Anthropic measuring Anthropic. The automation index rests on a task map the company drew of its own work and a rating scale the company defined; the block rate is the company's monitor counting its own interventions; the safety-compute share depends on which runs get labelled safety, by the people running them. No external party has verified any of it. This is precisely the failure we named when OpenAI published its disclosure framework four days ago — a process with no external reviewer grades its own homework — and the symmetry is uncomfortable: OpenAI built the process and skipped the reviewer, Anthropic committed to the reviewer and then published numbers ahead of them.

The timing is the sharp part. On 12 September Anthropic promised METR and others permanent employee-level access so outsiders could verify its safety claims. On 17 September it published the numbers those outsiders exist to check, with the evaluators, by its own account, still being set up. That is not bad faith — you publish what you have when you have it, and waiting would have been its own criticism. But it does mean the first pass of the new transparency looks exactly like the old transparency, and the difference between them is entirely in a future tense.

What would settle it is narrow and checkable: a version of these three numbers where the task map, the rating scale and the compute labels were audited by someone who does not work at Anthropic, published with the disagreements intact. Until that exists, 26% is a claim about a definition as much as about a capability — and the definition is the part nobody outside has seen.

What to watch

Whether the next edition carries an external auditor's sign-off or dissent, whether the automation index is published often enough to form a curve rather than a point, and whether any other lab publishes a comparable figure — which is when the definitions will have to be argued in public.