Anthropic reveals that 26% of its AI development work is now led by Claude

Summary

Anthropic's internal think tank reported that 26% of AI research and development tasks are now at the 'AI-first' (AL4) level. This ratio was below 1% in February. While humans oversee these tasks, collaborative AI usage is present in over 90% of them.

Why it matters

This provides one of the first concrete, quantitative snapshots of recursive improvement in a frontier lab. The cycle of 'AI building AI,' often hypothetical, is becoming visible in operational terms. The jump from under 1% to 26% in six months indicates that AI-assisted development speed is increasing much faster than most external observers model. This has direct implications for compute demand, safety oversight, and the competitive timeline among frontier labs.

Highlights

• Methodology: Claude classified approximately 15,000 tasks according to Epoch AI's AL0–AL5 scale by reading sampled staff Slack messages and documents. The AL4 level was measured at 26%, while the AL5 level was zero.
• Approximately 30,000 AI agents are operating concurrently on Anthropic's most heavily used infrastructure. The monitoring system blocked only 0.002% of over 1 billion decisions in August.
• According to a conservative estimate, about 6% of research and development compute resources within a week are allocated to safety work. Anthropic states it will continue to publish these metrics.

Critical take

The entire measurement is self-administered. Anthropic selected the sampling frame, had its own model, Claude, perform the classification, and published the results on its own timeline. Consequently, the 26% figure is as much a public relations product as it is a benchmark. The near-zero blocking rate (1 in 47,000) is presented as a reassuring indicator without independent audit or verification of what those 4,700 blocked actions actually were.