Anthropic Releases Metrics to Track Frontier AI Development Inside Labs

Anthropic published an overview of measurements for understanding the pace of AI development inside frontier AI labs. The company shared three distinct metrics covering how AI models build future versions of themselves, how agent actions are overseen, and how compute resources are distributed.

To track automated research and development, Anthropic created the Anthropic R&D Automation Index using an automation rating scale from AL0 to AL5. As of August 2026, the model Claude is not operating fully autonomously for any measured subset of AI R&D work. However, Claude leads 26% of Anthropic's AI R&D work, and the share of work at or above the level where AI collaborates is above 90%.

For agent oversight, Anthropic reported on approximately 30,000 agents performing research and engineering work on its most-used internal platform during August 2026. Online monitors achieved 100% coverage by checking actions before execution, with automated review happening before action and human review of blocked actions occurring within one week. Over a billion research and engineering decisions were analyzed over August 2026, finding that 0.002% of them were blocked. Offline monitors also achieved 100% coverage by ingesting all actions after completion, flagging roughly one to two transcripts in every thousand for further review, with approximately 50 high-priority flags per week escalated to human review.