OpenAI reports higher success rates for autonomous research agents
The company says agents completed 46% of research tasks requiring two to four hours of human work, but performance varied sharply by assignment.

OpenAI says its coding agents can independently finish 46% of research assignments that would require two to four hours of human work, compared with 23% in March. For assignments requiring four to eight hours, the company reported a success rate of 35%, versus 13% previously.
The findings appear in OpenAI’s second report examining how artificial intelligence accelerates its internal research. The company cautioned that researchers may have adjusted the assignments they delegated as agents improved, meaning the results reflect both model capabilities and changes in task selection.
Performance differed substantially across research activities. OpenAI reported independent completion rates of 79% for monitoring training runs and 72% for technical questions. The figures were 25% for inference and serving code, and 29% for evaluation and training environments.
Agent usage also increased. Valued at API list prices, median daily consumption per researcher reached about US$2,400 in October, compared with US$1,500 in September. At the 90th percentile, it reached about US$23,000 a day, versus US$11,000 in September.
Despite those gains, researchers interviewed by OpenAI said agents needed close supervision and remained weak at judging research ideas, code, data and writing. Post-training and safety researchers attributed all the main ideas behind GPT-6 Astra and GPT-6.2 Astra to people, while saying models helped implement them.
OpenAI called for laboratories to be required to monitor and disclose indications that AI is accelerating the development of future AI systems. It said that position did not mean it intended to pursue full self-improvement.




