Anthropic Says Claude Now Leads 26% of Its Own AI R&D Work, Up From Under 1% in February

Anthropic just published exactly the kind of number most companies refuse to share about their own AI rollout: how much of the work is actually being done by the AI, not the humans supervising it. Using an internal “R&D Automation Index” that scores tasks on a six-level automation scale weighted by person-time, Anthropic found Claude now leads 26% of the company’s own AI research and development work — up from under 1% in February 2026 — with more than 90% of that work reaching at least an “AI collaborates” level. No fully autonomous R&D operations exist yet, but the trajectory is the headline, not the current state.

The company paired that with two other measurements most labs keep internal: roughly 30,000 agents run simultaneously on Anthropic’s own platform under 100% automated monitoring coverage, with only 0.002% of decisions blocked and one to two transcripts per thousand flagged for human review — and a compute snapshot showing just 6% of total AI R&D compute, and 12% of AI-driven R&D compute specifically, going to safety research. Anthropic frames the release around a stated principle: “we should do everything possible to minimize the gap between what frontier labs know and what the public knows.” That follows a pattern the company set two months earlier with Inviting hard questions, a public initiative committing to track and report its own follow-through on AI’s hardest open questions rather than only publishing capability announcements.

For anyone advising a company on its own AI governance, the useful move here isn’t citing Anthropic’s specific percentages — it’s copying the instrument. Most enterprises can’t say what share of a given function AI actually leads versus assists, or what fraction of agent decisions get caught by monitoring versus a human. Anthropic just showed what it looks like to measure that internally and publish it anyway.