OpenAI Releases 722 Math Papers From Advanced Model

OpenAI has made 722 mathematical manuscripts, organized into 372 families, publicly available on GitHub under the Apache-2.0 license. These papers were generated by an unreleased internal frontier model that OpenAI claims is significantly more capable than its GPT-6 Astra system. This is the same underlying model responsible for the company's controversial Navier-Stokes singularity claim in September. While the release includes preprints, LaTeX sources, and some Lean formalizations, OpenAI has not provided API access, model weights, or inference capabilities.

To generate these results, OpenAI presented the model with roughly 4,000 problems. The generation process was largely fixed, with each successful result requiring an average compute allocation equivalent to three hours of ChatGPT Pro "thinking" time. The repository features 10 abridged reasoning trace summaries detailing the model's approach to complex problems, such as the irrationality exponent of pi, Kaplansky's direct-finiteness conjecture in characteristic two, the Mézard-Parisi formula, and quasipolynomial bounds for arithmetic progressions.

To address previous skepticism, OpenAI is routing the review process through a newly formed Advisory Group on Mathematics and AI at the Institute for Advanced Study. Practitioners can download the Lean files to mechanically verify the proofs, though formal coverage remains incomplete. For unformalized papers, mathematicians must conduct traditional peer reviews to check assumptions and ensure the proofs are valid.

For researchers and AI developers, this release provides a concrete framework for publishing AI-generated scientific discoveries. It establishes a protocol combining versioned preprints, BibTeX citations, and partial machine verification. If a significant portion of these papers survives expert scrutiny, it could radically accelerate the rate at which new mathematical literature is produced and verified.