OpenAI released 722 manuscripts describing mathematical results from an unreleased model, while mathematicians assess the work and its research practices.
OpenAI has published a collection of 722 mathematics manuscripts describing results generated by an unreleased frontier model. The papers cover 372 result families, and the Advisory Group on Mathematics and Artificial Intelligence (AGMAI) characterized the release as containing solutions to hundreds of open questions. The Verge reported the publication on October 6, 2026.
The materials include summaries of the model’s reasoning, estimates of computing used and counts of the problems attempted. OpenAI says the average result required computing equivalent to three hours of ChatGPT Pro thinking. The company released the papers through a GitHub repository and included procedures for revisions and citations, according to its description reported by The Verge.
The publication follows an earlier statement from OpenAI in September that its model had resolved more than 100 long-standing open problems across most areas of mathematics. The new collection provides a larger set of manuscripts and additional process details, but the results still require scrutiny by mathematicians. Their validity, significance and relationship to existing work will take time to assess.
That review is unfolding amid debate over how AI companies present mathematical discoveries and credit human contributions. AGMAI, an independent group of mathematicians formed to advise on communicating the results, has recommended prompt publication through established academic channels where possible, alongside disclosure of details such as the model, prompts and computing costs. It has also warned labs against using mathematical findings primarily as model promotion.
OpenAI said it is continuing to consider other community-hosted publication options that meet the group’s guidance and plans to improve the citations, exposition and presentation of future releases. The results add to a growing body of mathematical work attributed to AI systems, including efforts from rival labs, that researchers are still evaluating.