Anthropic's R&D Automation Index

AI & tech · Heat 1.0 / 10 · 2 shows · 11m airtime

Anthropic proposed a three-axis measurement framework for AI development pace on September 18, including an R&D Automation Index, agent oversight metrics, and compute allocation. It reports Claude leads 26% of R&D work and collaborates on over 90%, up from 1% in March before Mythos and 12% in May, with around 30,000 agents doing research and engineering work at any time and a 0.002% escalation rate. Arun Rao has called for standard measures across all labs, reported publicly weekly to monthly. Whether other labs adopt shared metrics remains open.

How 2 podcasts covered it

1. The AI Daily Brief, 2026-09-18 (2:46 to 5:20)
   Walks through Anthropic's three-axis framework
   Walks through Anthropic's proposed three-axis measurement framework: the R&D Automation Index, agent oversight metrics, and compute allocation. Reports Claude leads 26% of R&D work, 30,000 agents active at any time, and a 0.002% escalation rate.
   New angle: Anthropic says Claude leads 26% of its R&D work and 30,000 agents run at any time.
2. The AI Daily Brief, 2026-09-18 (5:49 to 6:37)
   Surveys outside reaction to Anthropic's metrics
   Surveys outside reaction to Anthropic's proposed metrics: safety researcher Jeffrey Ledish welcomes the transparency, Prime Intellect's Elie Bukhoush wants OpenAI's per-task usage breakdown combined with Anthropic's automation levels, and Arun Rao calls for standard measures across all labs reported weekly to monthly.
3. Big Technology Podcast, 2026-10-14 (55:47 to 1:03:47)
   Argues recursive self-improvement is real
   Roose uses Anthropic's R&D Automation Index to argue recursive self-improvement is real and dangerous, saying Claude leads 26% of R&D projects and could reach 50, 75 or 100 percent. The figure is already on the record; the rest is his own argument and prediction.