Claude Now Builds 26% of the Next Claude
Anthropic disclosed (Sept 17) that Claude now leads 26% of the company's AI research and development end-to-end, up from under 1% in February, with more than 90% of R&D involving Claude as at least a collaborator, roughly 30,000 AI agents running concurrently on its internal platform, and its safety monitor blocking about 1 in 47,000 of over a billion internal decisions — published as CEO Dario Amodei publicly calls for the industry to slow down.
Receipts
- primaryhttps://www.anthropic.com/institute/measuring-pace-of-ai-development
- alsohttps://abcnews.com/US/wireStory/anthropic-model-claude-helping-build-version-136547096
- alsohttps://www.nbcnews.com/tech/tech-news/anthropic-says-model-claude-helping-build-next-version-rcna598494
- alsohttps://thehill.com/policy/technology/6098297-anthropic-claude-leads-ai-development/
- alsohttps://www.anthropic.com/institute/recursive-self-improvement
Flagged:All figures (26% up from <1% in February, 90%+ collaborative R&D, ~30,000 concurrent internal agents, ~1-in-47,000 safety-monitor block rate across 1B+ decisions) are Anthropic's own disclosed metrics, published 2026-09-17 alongside a companion essay ('When AI builds itself') on recursive self-improvement; corroborated without discrepancy by ABC News (AP wire), NBC News and The Hill. A separate figure in the same disclosure — more than 80% of code merged into Anthropic's production codebase attributable to Claude as of May 2026 — measures a different thing (code authorship, not R&D task leadership) and was deliberately left off the card's three stats to avoid blending two distinct metrics into one story; it appears in the fbLinkedIn/reel body copy only, clearly separated. No lab, Anthropic included, reports any R&D task where its model operates fully autonomously without human supervision — the card and every caption preserve that qualifier rather than overstating 'end-to-end' as 'unsupervised.' Anthropic brand color #D97757 reused from the prior anthropic-ai-malware-rebuild-loop card, verified against the company's own logomark at the time.