A week after Dario Amodei’s essay argued the frontier should pace itself because models now build the next generation of models, Anthropic shipped the measurements. Its new “measuring the pace of AI development” post — built on Epoch AI’s automation ladder — puts Claude at or above the “AI collaborates” level in more than 90% of the lab’s R&D work, with roughly 30,000 research and engineering agents running at any one time, and the lab’s own monitors blocking about one decision in 47,000. Per Reuters’ read of the same material, Claude now leads roughly a quarter of the work on Anthropic’s next-generation models. The claim at the center of the essay is no longer argument; it is a lab-reported number.
For the operator, the self-building rate stops being an inference and becomes an input to roadmap risk: plan against the cadence the vendor measures, not the preview dates it prints, and expect the countdown-follow-up rule to keep paying rent. The same evening put the offense side of that automation on the table — per WSJ and The Information, bug-hunters used Claude to break into OpenAI. So the read is symmetrical: Claude is the thing building the next Claude, and tonight it was also the tool attackers reached for against the incumbent. The take that survives both is the one this column keeps circling: when the lab whose CEO wrote the pace-limit essay reports its own models leading a quarter of the build, the only clock left that has to reconcile a speed limit with a capital race is the S-1 that is still unfiled.
Source: anthropic.com; quarter-figure reported by reuters.com, apnews.com, washingtonpost.com; security beat per wsj.com and theinformation.com