Anthropic discloses internal metrics: 30,000 agents at work, 26% of R&D Claude-led

AI寒武纪 · wechat · 2026-09-18

Anthropic has for the first time published three core internal metrics to establish an industry transparency baseline.

R&D automation index (EpochAI-style AL0-AL5): as of August 2026, fully autonomous (AL5) development by Claude is 0%; Claude-led (AL4) work is 26%; deep AI collaboration and above exceeds 90%. About 30,000 agents run simultaneously on internal platforms, even delegating tasks to each other.

Real-time agent monitoring: two monitoring layers cover 100% of agent actions. In August, the online system analyzed over 1 billion agent decisions and blocked 0.002% (about 1 in 47,000), all human-reviewed within a week; offline monitoring flags 100k records weekly, with 50 top-priority ones escalated to safety staff.

Compute allocation: sampling one week in mid-July, safety research received only 6% of all AI R&D compute and 12% of compute dedicated to AI self-directed development — a conservative estimate that excludes safety classifier runtime.

Anthropic plans to admit independent third-party institutions with the same access as internal risk teams, hoping the industry adopts hard constraints like minimum safety-compute ratios and caps on agent compute.

Original post →

Anthropic discloses internal metrics: 30,000 agents at work, 26% of R&D Claude-led

More from Companies & People