OpenAI Researchers Spend $600 a Day on Coding Agents — 3.1× Human Labor
What shipped
OpenAI published Research acceleration: The view inside OpenAI this morning — its first hard-data snapshot of how coding agents reshaped work inside the research org. Buried in the disclosure: the company says it hit the goal announced last fall of an automated research intern by September 2026, with an automated AI researcher targeted for March 2028.
The receipts
Median researcher: more than $600/day of agent inference at API prices as of mid-August, up from "modest amounts" in January. 90th percentile: over $7,000/day. Agent runtime passed total human labor in June; the org now runs 3.1 agent-workdays per human workday. Experiments per active experimenter hit an all-time high in August since tracking began in Jan 2025. Success rates on delegated tasks rose from January to July — but over half of successful 4–8 hour tasks still required at least one human intervention.
The number that should worry policy people
On Aug 7, after evidence that Astra reached critical cyber capability, OpenAI restricted it. Astra-class GPU allocation fell 59.2% in a week — and allocation to other model classes rose 17.2%, offsetting about 85% of the decline. Controls on one model class didn't reduce compute demand. They redistributed it. That's the first public quantification I've seen of compute rerouting under model-specific restrictions, and it's a direct receipt against any safety policy that treats a model-class cap as a spend cap.
Why a builder cares
The per-researcher inference bill is now a named budget line: $600/day is roughly $18k/month per person — about 90× a $200/mo developer plan. If you're not instrumenting agent spend per employee, you're blind to the metric frontier labs actually track. And watch quiet support channels: OpenAI's internal office hours are dying because agents troubleshoot infrastructure, with one team canceling sessions entirely. Nobody games attendance numbers — that's the cleanest agentic-adoption signal you'll find.