Anthropic: Claude Now Leads 26% of Work Building Its Next Models
Reuters: AI collaborated on 90%+ of Anthropic R&D by August — not fully autonomous, with action screening that blocked about 1 in 47,000 decisions.
Anthropic told Reuters that Claude now “leads” 26% of the company’s own AI research and development work — up from 1% in March on an Epoch AI scale — and that AI collaborated with humans on more than 90% of research work as of August.
The company will publish these measures regularly so outsiders can track how fast AI is helping build the next generation of models. Claude is not fully autonomous in any measured part of the work.
What the numbers say
- ~30,000 AI agents on Anthropic’s main internal platform at once in August
- Every agent action screened before execution; ~1 in 47,000 decisions blocked that month (over a billion decisions)
- ~6% of AI-research compute went to safety in a July sample week (12% for research done by AI itself); conservative accounting when work advances both safety and capability
Operator takeaway
This is a process story, not a singularity headline. Leading labs are putting agents on real R&D throughput and putting gates on every action. Same leave/augment/automate map: automate draft and exploratory work, leave consequential commits behind screening, measure contribution honestly. If your shop can’t say what % of work AI leads — and what gets blocked — you’re flying without instruments.
Source: Reuters — Anthropic says Claude now leads a quarter of work building its next AI models (Aditya Soni).
