OpenAI has released new internal data showing that it has met the first of two recursive self-improvement goals it set last year: building an "AI research intern" by September 2026.
Under that roadmap, the company set two targets: an "AI research intern" by September 2026, followed by a fully autonomous AI researcher by March 2028. OpenAI now says the first milestone has been reached.
Agents are taking on parts of research workflows
According to OpenAI, once human researchers provide a clearly defined research problem, agents can independently complete some tasks that previously took skilled researchers several days. Researchers are also assigning work to agents more often.
Agent runtime reached 24.8 hours for every 8 human work hours
By mid-August, measured across the research department as a whole, every 8 hours of human work corresponded to about 24.8 hours of combined agent runtime, equal to roughly 3.1 agent workdays. OpenAI said this is a runtime measure, not a direct reading of research efficiency, and it should not be interpreted as a 3.1x increase in output.
Before June this year, total agent runtime was still below the human total. That has now changed.
Tasks are becoming longer and more complex
At the start of the year, agents were mainly used for research code and infrastructure code. They have now begun handling longer and more complex work, including experiment monitoring.
OpenAI also said that in August, the number of experiments run per active experimenter reached the highest level since the company began tracking the metric.
Humans still set research direction
Even with that progress, research direction is still decided by humans. Agents rarely handle high-level research planning. Over the past six months, more than half of successfully completed tasks lasting 4 to 8 hours still needed at least one human intervention.
OpenAI's next target is to turn this "intern" into an autonomous AI researcher capable of completing a more complete range of research work by March 2028.