OpenAI reports uneven gains from coding agents in its research work

OpenAI published an internal research-acceleration report on November 18, describing rising coding-agent use alongside substantial differences in autonomous task performance.
The company estimates that success without human intervention on research tasks requiring two to four hours of human work rose from 23% in March to 46% in October. For four-to-eight-hour tasks, it rose from 13% to 35%.
Results varied sharply by activity: estimated success was 79% for monitoring training runs but 25% for writing inference and serving-stack code. OpenAI cautions that changing task-selection practices, not just better models, may influence those trends.
At API list prices, the median researcher’s daily agent usage reached $2,400 in October. That figure values usage at published prices; it is not a statement of OpenAI’s actual internal computing cost.
Interviews also described agents wasting effort, producing low-quality code and making invalid analysis assumptions. Humans remained closely involved.
The report supplies company-generated evidence about AI-assisted research, while showing why more agent activity should not automatically be equated with autonomous scientific progress.
Sources and citations
Revision history
Published .