Anthropic report: Claude leads 26% of its own AI R&D work

Anthropic's own measure says Claude now "leads" 26% of its AI research and development work, up from under 1% in February 2026. The company published those numbers in a piece on its website co-authored by Marina Favaro and Phillie Wright.

"We want to illuminate the pace of progress for the public," the authors write. "As the world considers pacing the frontier, we should do everything possible to minimize the gap between what frontier labs know and what the public knows."

The index itself, the Anthropic R&D Automation Index, rates tasks on a six-level scale from AL0 (no AI involvement) to AL5 (fully autonomous). AL3 is "collaborates"; AL4 is "leads." Over 90% of surveyed work reaches at least AL3; nothing yet reaches AL5. The methodology sampled 20% of staff weekly through July 2026 and produced roughly 15,000 tasks sorted into a 542-node hierarchy with 378 leaf categories. Human reviewers agreed exactly with the model's ratings 59% of the time, and within one level 97% of the time.

The oversight numbers describe a system running at scale. Roughly 30,000 agents work simultaneously on Anthropic's primary internal platform. "100% of these agents' actions pass through the online monitor before they are executed, usually within seconds," the piece states, and the blocker fires on about 0.002% of decisions, or one in 47,000. A separate offline monitor ingests every action after the fact and flags "roughly one to two transcripts in every thousand for further review," with about 50 cases a week escalated to humans.

On compute, the piece reports a July 13-20 snapshot: "About 6% of compute that went to AI R&D was allocated toward safety, and about 12% of compute that went to AI-driven AI R&D was allocated toward safety." Ambiguous workloads were classified as capabilities rather than safety by default.

The figures are self-reported and not externally audited. Two of the researchers in our tracker shared the piece.