Claude Now Does a Quarter of Anthropic's Work

Anthropic just put a number on a claim that used to be a speech.

On 17 September 2026 it published the first snapshot of an R&D Automation Index. As of August, Claude "leads" 26 percent of the company's own AI research and development. In February that share was under 1 percent. More than 90 percent of the work is now at least "collaborates" — Claude does large chunks under close human direction. Zero of the measured work is fully autonomous.

That last line is the one Anthropic wants you to read twice. Leads is not autopilot.

Image: Anthropic, Measurements for understanding the pace of AI development inside frontier labs, 17 September 2026.

What "leads" actually means

The scale comes from Epoch AI. AL0 is no AI. AL5 is no human in the loop. AL3 (collaborates) is a person staying in the problem: they bring logs, they decide when a surprise is worth a proper fix, they review the change. AL4 (leads) is a high-level prompt and a supervisor who reads the write-up at the end. Claude still does not ship the pipeline. The engineer does.

Anthropic's own example is a broken nightly data job. At AL4, Claude reads the logs, finds the failing stage, writes and tests the fix, and documents extra patches. It tags a human. The human still decides if it goes out tonight.

How they counted

They sampled 20 percent of staff in the model R&D loop, each week in July. A Claude research agent listed tasks from Slack and internal docs. That produced about 15,000 granular tasks, then a tree of 542 nodes and 378 leaves (eval-platform bugs, RL sandbox policy, serving postmortems). A frozen basket so next month's number is comparable.

An independent Claude judge then rated each category. Weights came from person-time, not from execs picking favorite teams. Model-versus-human exact agreement was 59 percent. Human-versus-human was 35 percent. Ratings were within one level 97 percent of the time. Anthropic admits the judge is its own model. That is the circularity. Third-party evaluators are the planned patch.

30,000 agents, and what actually gets stopped

On the main internal platform, about 30,000 research and engineering agents run at once. Online monitors see 100 percent of actions before they execute and blocked 0.002 percent of them in August — about 1 in 47,000 — across more than a billion decisions. Offline monitors flag 1 to 2 transcripts per thousand. About 100,000 flags a week; roughly 50 of those reach a human.

Agents get individual identities that survive model upgrades. They talk on an open channel, not private whispers, so one agent can treat another agent's claim as something to check. That is an architecture choice, not a safety slogan.

Where the chips went

For the week of 13–20 July, about 6 percent of AI R&D compute was classified as safety work, and about 12 percent of the AI-driven share. Anthropic calls both conservative: mixed work was counted as capabilities. It is one week, tagged by another Claude classifier. Useful as a template. Weak as a trend.

What to do with this if you run agents

Copy the four numbers, not the 26 percent press line: share of work agents lead, concurrent agent count, block/escalation rate, and safety compute share. A standards body will ask for those. So will a buyer. A dashboard you only publish when the number looks good is marketing.

This sits next to Amodei's call to slow the frontier. The same week, the lab showed why the curve is steep — and that AL5 is still empty. Watch the next snapshot. If "leads" keeps jumping and "autonomous" stays at zero, the story is supervision at scale. If AL5 lights up, the story changes.

Sources

Anthropic, Measurements for understanding the pace of AI development inside frontier labs, 17 September 2026. https://www.anthropic.com/institute/measuring-pace-of-ai-development

Reuters, Anthropic says Claude now leads a quarter of work building its next AI models, 17 September 2026. https://www.reuters.com/business/anthropic-says-claude-now-leads-quarter-work-building-its-next-ai-models-2026-09-17/