- Prototype ‘Anthropic R&D Automation Index’ catalogues ~15,000 R&D tasks and assigns automation levels AL0–AL5.
- Headline: Claude at AL4 (“leads”) for 26% of weighted R&D work; >90% of work at or above ‘AI collaborates’ (AL3).
- Anthropic notes Claude is not fully autonomous in any measured area and treats the index as a prototype open to independent verification.
- Anthropic describes coverage, review latency, and escalation rate as oversight metrics for agent actions.
- Example snapshot: ~30,000 R&D agents active at times; in one sample week ~6% of compute for AI R&D was safety work (12% for AI‑driven AI R&D).
- Company says actions are automatically checked before execution and plans embedded third‑party evaluators to verify practices.
- Anthropic frames the metrics as a way to track how close labs are to recursive self‑improvement.
- Company is urging other frontier labs to publish comparable measurements and to allow independent verification.
- Anthropic will publish these measures regularly as part of its public risk reporting.
Useful?