OpenAI's coding agents are doing more of its research work than ever, and the company says the shift is accelerating. In a new report published on November 18, OpenAI found that agents now complete nearly half of the research tasks that take two to four hours, while humans remain central to the work.

The numbers moved sharply between March and October. For tasks lasting two to four hours, agent success rose from 23% to 46%. For longer tasks of four to eight hours, it rose from 13% to 35%. The results vary by job: agents succeeded in 79% of tasks involving monitoring a training run, 35% of tasks involving launching a training run, and 25% of tasks involving writing code for inference and serving systems. Agents still often fail at work that needs long-term planning, and they can make costly mistakes.

Use is climbing just as fast as capability. The median researcher's daily agent use reached $2,400 in October at API list prices, up 62% from September, and agent workdays per researcher rose from 0.5 to 5.7 per day. At the same time, researchers are devoting more of their effort to Safety and Alignment, which saw the largest absolute increase of any area.

The report builds on a post OpenAI published in September and adds quantitative measures of how agents are changing research. Its central argument is that labs should track and report indicators of recursive self-improvement, where AI systems help build the next generation of AI systems, so the progress can be measured openly rather than guessed at.