An Anthropic institute note dated 17 September placed Claude at 26 percent of the lab's August AI research-and-development load on an Epoch AI scale, with a collaboration share above 90 percent and no fully autonomous slice. About 30,000 agents ran together on the busiest internal platform. A July sample week sent about 6 percent of that compute slice to safety. The percentages are the lab's own measurements.
Artificial Intelligence··Midday
Leads, not autonomy
Anthropic, in a 17 September institute note, put Claude at 26 percent of the lab's own AI research-and-development load for August, with a collaboration share above 90 percent, on an Epoch AI automation scale. Lead, labelled AL4 there, means finishing most of a job from a high-level prompt while a person still watches. No measured slice was called fully autonomous. Readers should treat every percentage here as Anthropic speaking in its own post, not as an audited outside ledger.[1]
Thirty thousand agents
The same note described about 30,000 agents running together on the busiest internal platform. A monitor, Anthropic said, reviews every action before execution. On more than a billion August choices, roughly one stop landed among 47000 actions, matching 0.002 percent. Those ratios are the lab counting itself. Readers should treat every percentage here as Anthropic speaking in its own post, not as an audited outside ledger.[1]
A July safety share
A 13-20 July window sent roughly 6 percent of the AI research compute slice to safety work. Mixed safety-and-capability jobs stayed outside that bucket. Outside verification is still a plan in the note, not a finished audit. The retained institute page is the only pool card for these August shares; nothing in this slot adds a second publisher's own count.[1]
