Anthropic’s Claude Now Leads 26% of Its Own AI Research and Development

Author

AI News Editorial

Published

2026-09-18 10:15

Anthropic published the first results of its prototype R&D Automation Index on September 17, revealing that Claude now leads 26% of the company’s own AI research and development work as of August 2026—up from under 1% in February. The measurement uses Epoch AI’s Automation Level scale, which ranges from AL0 (no AI involvement) to AL5 (fully autonomous with no human in the loop).

The index was built by randomly sampling 20% of staff across departments in July 2026, identifying approximately 15,000 granular R&D tasks from Slack and internal documents. These were organized into a 542-node hierarchy with 378 leaf nodes, with an independent Claude judge assigning automation levels weighted by person-time.

What the Numbers Mean

At AL4—“leads”—the model completes most of a task end-to-end from a high-level prompt while a human supervises. Work at AL3 (“collaborates”) or above now accounts for more than 90% of Anthropic’s research, meaning almost none of the company’s R&D happens without the model in the loop anymore.

Thirty thousand research and engineering agents operate simultaneously inside Anthropic. Online monitors cover 100% of agent actions before execution and block roughly 1 action in 47,000. Offline monitors flag about 100,000 transcripts per week, of which approximately 50 escalate to humans.

Zero percent at AL5—the fully autonomous level—is the line Anthropic wants read carefully. Fully autonomous AI R&D is exactly the scenario that Dario Amodei’s recent “pacing essay” warned about when arguing for a six-to-twelve-month slowdown.

Safety Compute: 6% of R&D

In the week of July 13-20, 6% of AI R&D compute and 12% of AI-driven AI R&D compute went to safety work—figures Anthropic calls deliberately conservative. This is the first time any lab has published such a ratio.

The disclosure comes just six days after Amodei publicly asked the industry to slow down. The timing is unlikely coincidental: the automation index provides the quantitative backing for why the pacing concern matters.

“If Claude leads a quarter of its own development today and the curve holds, the next index reads 50% by spring,” noted one observer. “That’s the number the botnet warning was actually about.”