OpenAI Says Its Coding Agents Now Work Three Days for Every Human Day
OpenAI published internal figures showing its research organisation uses 3.1 agent-workdays of coding-agent runtime for every human workday, and says it has hit its goal of an automated research intern by September.
OpenAI published a set of internal numbers yesterday about how much of its own research is now done by AI agents, and the headline figure is unusual for its specificity: 3.1 agent-workdays of coding-agent effort for every one workday of human labour across its research organisation. The company says it has reached its stated goal of fielding an “automated research intern” by September 2026.
The metric deserves a little unpacking, because it is doing a lot of work. OpenAI took the total runtime of its coding agents, converted it into standard eight-hour workdays, and compared that to human labour. Before June 2026 the combined agent runtime sat below total human hours. By mid-August it was past three to one. Runtime is not output, and an agent can spend eight hours going in circles, so this measures effort applied rather than problems solved. OpenAI also reports that its heaviest internal users spend more than $7,000 a day on tokens, which gives some sense of the scale involved. The company defines “research intern” carefully: a system that carries out well-defined tasks under human direction, including ones that would take a skilled researcher a few days. Not a scientist choosing its own agenda. OpenAI says it is targeting a genuinely automated AI researcher by March 2028.
Why publish this at all. Partly it is a recruiting and investor signal, and partly it is an answer to a question people keep asking labs: if these tools are as good as you say, are you using them yourselves? This is the first time a frontier lab has put a concrete internal ratio on the table rather than gesturing at productivity. Treat it as a self-report, because that is what it is, and because “agent-workdays” is a metric OpenAI invented and defined. Still, the underlying claim is checkable in a slower way: if research really is accelerating three-fold, it should show up in shipping cadence, and the past six weeks have been unusually busy across the whole industry.
What this means for you. If you write code, the practical lesson is not the ratio but the shape of the work. OpenAI describes agents handling bounded tasks with a clear definition of done, then handing results back for a human to judge. That is a workflow you can copy today with the tools you already have, and it is a good deal more productive than asking a model to do something vague. If you do not write code, the interesting bit is what “intern” implies: supervised, bounded, and reviewed. Everyone experimenting with agents at work is converging on the same shape, which is a real finding even if the number attached to it comes from a company with an obvious interest in the answer.
Sources
Source: https://openai.com/index/research-acceleration-view-inside-openai/
OpenAI's Chief Scientist Says No Lab Has Solved Alignment Well Enough to Keep Going This Fast
Jakub Pachocki published an essay called An Alien Mind on 6 September arguing that no lab has solved alignment and monitoring enough to keep scaling at maximum speed, and that he expects voluntary slowdowns.
