OpenAI says it has reached a goal it set last fall: building what it calls an automated research intern. By this term, the company means an AI system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled human researcher a few days. In a detailed announcement, the company says it hit this milestone in September, and that it is now working toward a more capable automated AI researcher by March 2028.
By mid-August, the median researcher was using agents daily, consuming more than 600 dollars per day of computing power at API prices, and the heaviest users topped 7,000 dollars per day. Measured in standard eight-hour workdays, the research organization now uses 3.1 agent-workdays of effort for every human workday, up from less than one earlier in the year.
More experiments, but humans still steer
Researchers are running more experiments than before, with August marking an all-time high since tracking began in January 2025. OpenAI notes this coincides with wider adoption of its Codex coding tool, though its available computing power also grew. Using a taxonomy of research work from Epoch AI, which sorts AI research into six phases from deciding what to work on through building, running, and communicating (see this summary of the taxonomy), the company found that agents now help across all phases, but high-level planning remains a minimal share. Agents still need close supervision: more than half of successful tasks estimated at four to eight hours of human effort required at least one human intervention.
OpenAI also describes safety steps, saying that after a recent security incident at Hugging Face it paused reinforcement learning training, a method that trains models by rewarding desired behavior, on its newest deployment-bound models while it hardened its research environments. The company says it does not yet know how to safely reach aligned, full recursive self-improvement, meaning AI systems that improve themselves, and argues labs should be required to publicly track such progress.
This matters because automating AI research could accelerate the entire field, for better or worse. The figures are self-reported and preliminary, so they show both how fast agentic tools are spreading and how much depends on outside scrutiny of such claims.