OpenAI has published a large collection of mathematical manuscripts and supporting proof materials produced by an unreleased internal frontier AI model, offering a new look at how advanced AI could contribute to research-level mathematics. The collection includes work at different stages of verification, so not every manuscript should yet be treated as a confirmed mathematical result.
The public collection contains 722 manuscripts organized into 372 result families. A family can contain a main result along with related arguments, consequences, companion papers or alternative proofs. OpenAI says the broader evaluation involved roughly 4,000 mathematical problems across multiple areas of research.
The model behind the vast majority of the manuscripts has not been publicly released or named. OpenAI describes it as an internal frontier model and says it is working toward a responsible release of the system that produced the research.
An important part of the release is the use of Lean, a programming language and proof assistant that allows mathematical proofs to be checked by computer. OpenAI has included Lean formalizations for many of the manuscripts, giving researchers computer-checkable proof artifacts for those results. However, not every manuscript currently has a Lean formalization.
OpenAI has also published 10 abridged summaries of the modelโs reasoning for selected results. The examples cover topics ranging from the irrationality exponent of ฯ and Mahler conjectures to theoretical computer science and mathematical physics.
According to OpenAI, each retained result used, on average, compute equivalent to roughly three hours of ChatGPT Pro thinking with the internal model. Across the evaluation, the model was presented with approximately 4,000 problems before the outputs were organized into the published manuscript collection.
The release follows OpenAIโs earlier work on difficult mathematical research problems, including its work on the NavierโStokes problem. These developments have drawn attention both to the growing capabilities of frontier AI systems and to questions about how AI-generated mathematical research should be checked and communicated.
OpenAI has also been consulting the independent Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study. The group advises on the review, significance and communication of emerging AI-generated mathematical results, as well as academic and professional standards for presenting this type of research.
Independent scrutiny remains essential. OpenAI notes that the manuscripts are at different stages of verification and that some unformalized results may contain errors. Even when a proof is formally checked, mathematicians still need to assess its significance, assumptions, relationship to earlier work and whether it genuinely advances the field. The development also fits a wider shift toward advanced AI reasoning, as frontier models take on longer and more complex scientific and technical tasks.
OpenAI has published the collection in a public GitHub repository with protocols for citations, corrections and revisions. The company says earlier public versions will remain accessible when manuscripts are updated, allowing researchers to track how individual results change during review.
The development also raises broader questions about the future of mathematical research. AI systems may increasingly help researchers explore possible solutions, test mathematical ideas and work through problems that would otherwise require significant human time.
The release also shows why expert mathematical judgment remains important. Producing a candidate proof is only part of research: mathematicians must determine whether a result is correct, genuinely new, significant and properly connected to existing mathematical literature.
OpenAI says it plans to support workshops, conferences and special programs focused on understanding major mathematical results produced by AI. The company is also working toward responsibly releasing the internal model behind the latest collection.
The significance of the 722 manuscripts will become clearer as mathematicians inspect, verify and compare them with existing research. If substantial portions of the collection are confirmed as correct, novel and important, the release could become a notable example of frontier AI moving beyond mathematical benchmarks and contributing directly to research-level mathematics. The release could become a notable example of how artificial intelligence research is expanding into advanced science.
