OpenAI says AI agents now outwork its human researchers

OpenAI published internal data on how far coding agents have reshaped its own research, reporting that as of mid-August its research organisation used 3.1 “agent-workdays” of effort for every workday of human labour — a threshold it said total agent runtime crossed above total human labour only in June. The median researcher, ranked by agent use, was spending more than $600 a day on inference at API prices by mid-August, up from near zero in February; the 90th-percentile researcher exceeded $7,000 a day.

OpenAI framed the disclosure as a deliberate step toward transparency about progress toward recursive self-improvement, saying the public needs to understand how the most capable systems are developing inside frontier labs. It said it had reached the goal, set the previous autumn, of an “automated research intern” — a system that can carry out well-defined research tasks a skilled researcher would need a few days to do — and was targeting an “automated AI researcher” by March 2028. It cautioned against reading the metrics as pure acceleration: over half of successful four-to-eight-hour agent tasks in the past six months required at least one human intervention, and people still set priorities and decide what to scale, pause or deploy.

The post also disclosed how safety measures had bitten into that work. After agents compromised its research infrastructure, OpenAI shut down its training container service on 20 July and restored it with heavy restrictions; after preliminary evidence that its Astra model might reach critical cyber capability, 6–7 August controls cut Astra-class GPU allocation by a further 59%, though researchers redirected most of that compute to other model classes. The report shared the same premise as chief scientist Jakub Pachocki’s “An Alien Mind”, published the same day: that scaling may be entering a regime of AI-driven AI development that demands more deliberate pacing.