Anthropic: Claude now leads 26% of its model R&D, up from under 1% in February
RealGeneKim · x · 2026-11-11
Anthropic published an overlooked post unveiling three measurement tools to track AI self-development inside frontier labs, plus an internal snapshot:
- Claude now leads (AL4) 26% of Anthropic's model R&D tasks, vs under 1% in February; AI collaborates on or leads over 90% of R&D (about 4% in Aug 2025)
- 30,000 AI agents do research and engineering at any one time, making 1B+ decisions a month; only 1 in 47,000 gets blocked
- Monitors flag 100,000 agent transcripts weekly; only 50 reach a human
- Claude agents read staff Slack and docs to index 15,000 R&D tasks
- Claude agreed with humans more often than humans agreed with each other (59% vs 35%)
- Agents keep identities across model upgrades and coordinate on an open message board to check each other's work
Anthropic also plans to embed independent third-party evaluators to verify safety practices and monitor these metrics.
More from AGI Musings
- iamtrask: Email is the 'dial-up' version of the AI internet — iamtrask · 2026-11-11
- AI didn't break education — education was measuring the wrong things all along — Yamapama · 2026-11-11
- BVSR: less context during ideation makes AI scientists more creative, Stanford study — james_y_zou · 2026-11-11
- repligate notes Christiano was labeled an "OpenAI board member" — repligate · 2026-11-11
- Cognitive scientists find AI generates ideas faster, but human teams stay more creative — CurieuxExplorer · 2026-11-11
- Paul Christiano warns AI could kill everyone next year, wording sparks debate — repligate · 2026-11-11