Claude leads 26% of Anthropic's AI research, but no measured part of it is fully autonomous
Anthropic also counted the agents inside its own systems: about 30,000 running at a time, and a monitor that blocked one action in 47,000.
Anthropic published three internal measurements of its own AI development on September 17, 2026. The first one puts a number on how much of the company's AI research Claude now runs, and the number is 26%.
That's the reading for August 2026, and "leads" has a precise meaning here. Claude takes most of a task end to end from a high-level prompt, and the engineer supervises rather than steers.
- Published
- September 17, 2026
- Claude leads
- 26% of Anthropic's AI R&D work
- AI collaborates or above
- above 90%
- Agents at once
- approximately 30,000
- Actions blocked
- 0.002%, about 1 in 47,000
- Compute to safety
- about 6% of AI R&D compute
What "leads" means here
The scale comes from Epoch AI. It runs from AL0, no AI involvement at all, up to AL5, fully autonomous with no human in the loop, and "leads" is AL4. Anthropic calls its own version the Anthropic R&D Automation Index.
Nothing they rated has reached AL5. (The footnotes work the levels through on a broken nightly data pipeline: at AL4 Claude finds the failing stage, writes and tests the fix, then tags an engineer to decide whether it ships tonight.)
Claude is not operating fully autonomously for any measured subset of AI R&D work.
One level down, at "AI collaborates", Claude does large chunks of work under close human direction. More than 90% of the work sits at or above that line.
Thirty thousand agents, one blocked action in 47,000
The second measurement counts agents. In August 2026 roughly 30,000 of them were doing research and engineering work at Anthropic at any one time on its most-used internal platform, and every action they take goes through a monitor before it runs. Usually within seconds.
Out of more than a billion decisions that month, 0.002% got blocked, which is about 1 in 47,000. A second monitor reads the transcripts after the fact and flags one or two in every thousand; roughly 50 a week reach a human.
The two monitors are built for different kinds of trouble. The online one catches harms that happen fast and can't be undone (the example they give is an agent copying model weights out of Anthropic's systems), and a human looks at any blocked action within a week. The offline one is for the slow kind, like signs of misalignment.
In our monitoring data to date, individual agents rarely misbehave. But when there are millions or billions of agents operating in the economy, even rare events can happen regularly.
Six percent of the compute
For the week of July 13 to July 20, about 6% of the compute Anthropic spent on AI R&D went to safety work. Of the compute that went to AI-driven AI R&D, the safety share was about 12%.
Anthropic calls those estimates deliberately conservative. A token that advanced capabilities as much as it advanced safety wasn't counted as safety at all, and safeguards classifiers sit outside the figures entirely. Granted, it's one week, which the company says is enough to show the measurement can be made and not enough to show a trend.
To be fair to the 6%, safety research doesn't eat compute the way a frontier training run does, and the post says so. Anthropic's case for it is that it compares like with like, across developers and over time.
The number that surprised me
It's in the appendix. Anthropic checked its judge model against the staff who own each area of work, and those staff rated their own areas without seeing what the models had gathered. The model matched a human rating exactly 59% of the time. Two humans matched each other 35% of the time.
So the people doing the work disagree with each other about how automated it is more often than the model disagrees with them. Ratings landed within one level of each other 97% of the time, and Anthropic says there's real room for disagreement on borderline cases, such as where "AI collaborates" ends and "AI leads" begins. I'd want that 35% printed next to the 26% every time this index gets quoted.
we should do everything possible to minimize the gap between what frontier labs know and what the public knows
Who checks the checkers
All of this is Anthropic measuring Anthropic with Anthropic's own models, and the post is direct about the cost: the judge model could make the same kinds of errors as the model it's checking. Comparing one lab against another hits a second wall: there's no common methodology.
The remedy Anthropic names is the one Dario Amodei committed the company to in his September 12 essay: independent third-party evaluators from several organizations, working inside Anthropic with access comparable to what the internal risk assessment teams get, to verify safety practices and monitor metrics like these. Anthropic doesn't say when they arrive.
The 26% is a reading of August 2026 against a basket of tasks frozen in July 2026 (about 15,000 of them, sorted into a tree of 542 nodes). Anthropic says it plans to rebuild that basket periodically and re-version the numbers.