OpenAI said on Sept. 6, 2026, that by its own measurements it has reached its goal of an automated research intern: a system able to carry out well-defined research tasks under human direction, including tasks a skilled researcher would need a few days to do. The work is done by AI agents, AI that carries out tasks on its own instead of only answering questions. OpenAI is claiming that its agents now put in more research time than its people, measured in workdays, while still needing people to steer the hard tasks.
By mid-August, OpenAI said, its research organization was using 3.1 workdays of agent effort, counted in standard eight-hour days, for every workday its people put in. Before June 2026, the agents' total running time was still below the time people worked. Among researchers who use agents every day, the median one was running more than $600 a day of model use, priced at what OpenAI charges outside developers. At the 90th percentile, a research user was running more than $7,000 a day.
OpenAI describes AI research as a loop with many steps. Researchers design improvements, write tests to judge models and the systems to try ideas at scale, catch bugs as well as unsafe behavior or behavior that strays from what people intend, and fold winning ideas into a core training run. To sort the agents' work, OpenAI used a six-part breakdown of AI research from the research group Epoch AI: Decide, Design, Build, Run, Analyze and Communicate. Agent work grew in every part from January to August 2026, but high-level planning stayed a tiny share of what the agents produced.
Using an AI system to classify the tasks, OpenAI said success rates generally rose from January to July across several levels of difficulty. But the agents still need a lot of human steering as tasks get harder, the company said: in the last six months, more than half of the successful tasks that would take a person four to eight hours needed at least one human step-in. OpenAI's report does not say how reliable that classifying system is. People still set research priorities, judge which ideas to pursue, and decide whether to scale up, pause or roll out systems.
Experiments per active researcher hit their highest level in August 2026 since tracking began in January 2025, OpenAI said. It said the rise goes along with wider use of Codex, its coding agent, though its available computing power has also grown significantly. The report's own methods section warns that such activity measures are relatively easy to collect but hard to interpret, because how they relate to research progress is uncertain, and that its coding-agent figures cover most but not all agent use. OpenAI has set March 2028 as its target for an automated AI researcher.
OpenAI's report cites no outside review of the intern milestone, so the ratios are the company's own figures. It is also unclear whether more agent running time and more experiments will track real research progress once the steps that are hardest to automate become the bottleneck, and how much human steering the agents will need as tasks grow longer.