OpenAI has released 722 manuscripts that it says contain solutions to hundreds of open questions in mathematics, a sweeping batch of work produced by an unreleased frontier model. The papers are grouped into 372 result families, giving the release a scale that is unusual even for a field accustomed to slow, careful progress.
The timing matters because the company is now putting concrete details in front of mathematicians after weeks of anticipation. In September, OpenAI said the model had resolved more than 100 long-standing open problems across most areas of mathematics, and the new repository adds summaries of the model’s reasoning, estimates of the compute used and counts of how many problems were attempted. OpenAI said the average result used the equivalent of three hours of ChatGPT Pro thinking.
That scale is part of why the release is landing so hard. For mathematicians, the issue is no longer whether AI systems can occasionally find new paths through hard problems, but how those results are checked, cited and absorbed into the literature. OpenAI said it is publishing the work in a GitHub repository with protocols for paper revisions and citations, and it said it is still exploring other community-hosted alternatives that fit the committee’s guidelines.
The release has also sharpened an argument that has been building around OpenAI mathematics work more broadly. AGMAI, in its first recommendations published in late September, urged AI labs to release mathematical results promptly and through established academic channels where possible, while also asking them to disclose the model name, prompts and compute costs. It also warned companies not to turn mathematical results into marketing vehicles for their models, a line that speaks directly to the unease some mathematicians have expressed as AI-generated results have arrived in ever-larger batches.
This is now bigger than a single paper or even a single model. OpenAI’s results add to a growing body of mathematical work from OpenAI and rival labs like Anthropic, and the field is still processing results tied to a Millennium Prize problem. The next test is whether the manuscripts can be taken up as serious mathematics rather than treated as a product announcement, because the answer will shape how quickly AI-generated proofs become part of the discipline rather than a parallel stream beside it.
