On October 6, OpenAI published a batch of mathematical research produced by an internal frontier model on GitHub, in a repository called openai/math, released under the Apache 2.0 license. The repository contains 722 manuscripts, grouped into 372 research "families" — different proofs, generalizations, and revisions of the same problem collected together.
According to OpenAI, the model was assigned about 4,000 problems over the course of the project. Each accepted result consumed, on average, roughly the amount of compute equivalent to three hours of ChatGPT Pro's extended thinking.
Which fields are covered
The manuscripts span number theory, computational complexity, mathematical physics, geometry, and theoretical computer science. Examples cited in media coverage include problems related to irrationality measures of π, several NP-hard problems, the Mahler conjecture, arithmetic progressions, and free group factors, as well as the quantum Heisenberg ferromagnet and the relativistic Vlasov-Maxwell equations.
Many of the proofs come with a formal version written in Lean, a programming language that lets a computer check a proof's logic step by step; passing formalization means there are no gaps in the reasoning chain. Not every manuscript has a matching Lean file, though, and OpenAI was blunt about it in its notes:
"some of the unformalized results could have issues"
In other words, results that haven't been formalized may still contain errors. The company has committed to fixing any mistakes that are found and to gradually adding more formal proofs. The repository also comes with a paper-revision process and citation guidelines, so researchers can track how many times a given result has been revised and which version to cite.
Connecting the dots
This release comes more than four months after OpenAI's last major math milestone. Back in May, an OpenAI model resolved a roughly 80-year-old open Erdős problem — a single breakthrough reported one problem at a time. By October, that had turned into hundreds of manuscripts dropped into a repository at once; the manner of the release became part of the story itself.
The format changed too. Before this release, OpenAI consulted an independent "Mathematics and AI Advisory Group" convened by Princeton's Institute for Advanced Study (IAS). That group published a set of recommendations on September 29 — drawing on more than 600 pieces of feedback from the mathematics community — calling for AI-generated math results to disclose the model's name, the prompts used, a summary of its reasoning, the compute cost, and the model's failure rate on similar problems.
Measured against that list, OpenAI this time provided a compute figure (an average of three hours of Pro-level thinking), a total problem count (about 4,000), and a large batch of Lean proofs; the model itself remains undisclosed, with the announcement referring only to an "internal model" and saying the company is working toward releasing it responsibly. How many of those roughly 4,000 problems went unsolved, and which were abandoned partway through, isn't clear from the public materials.
Mathematicians have long worried about this kind of output volume. Terence Tao has publicly raised the issue twice, in August and September, arguing that AI could let the production of proofs far outpace humans' ability to review them, with good hard problems getting "mined" quickly. With 722 manuscripts released at once, who ends up carrying the review burden is a question nobody has answered yet.
What comes next
OpenAI says it will fund related workshops and dedicated programs to help the mathematics community understand and absorb the AI-generated results. The company hasn't laid out a concrete plan for whether the manuscripts could make their way into formal journals or how authorship would be handled, and outside mathematicians' manuscript-by-manuscript review results haven't been published either.
Sources: OpenAI's official blog, Interesting Engineering, CocoLoop, Unite.AI; manuscript counts, problem totals, and compute figures follow OpenAI's public statements, and the advisory group's recommendations follow the IAS group's published document.