Anthropic says Claude now leads 26% of its internal AI R&D, up from <1% in February

机器之心 · wechat · 2026-09-18

Anthropic has published the first quantified look at its internal R&D automation: as of August, 26% of its AI R&D work reached AL4 ("AI leads") on EpochAI's AL0-AL5 scale, up from under 1% in February, while over 90% of work is at least at AL3 ("AI collaborates"). No measured work has reached fully autonomous AL5.

Key facts:

The number comes from an "RD Automation Index": Claude analyzed 15,000 concrete R&D tasks sampled from employee Slack/docs, organized them into a 542-node work tree, and rated each via an independent Claude judge.
About 30,000 agents are running research and engineering on Anthropic's internal agent platform at any moment, making over 1 billion decisions in August alone; only 0.002% were blocked by an always-on monitor, which checks 100% of agent actions before execution.
Agents have persistent identities, delegate tasks to each other, and communicate over a shared messaging system designed so claims from other agents are treated as unverified.
6% of AI-R&D compute goes to safety research; among AI-driven R&D, that share is 12%.
Earlier disclosures: over 80% of merged code in May was attributed to Claude, and typical engineer code-merge volume is 8x the 2024 level.

The post frames this as a direct signal toward recursive self-improvement (RSI): if the trend holds, AI-led work could approach 80% by end of 2026.

Original post →

Anthropic says Claude now leads 26% of its internal AI R&D, up from <1% in February

More from AGI Musings