OpenAI dumps 722 math papers written by an unreleased model, and says some could be wrong

On Tuesday, Oct. 6, 2026, OpenAI published a public collection of 722 math manuscripts produced by an internal model it has not released, along with machine-checked Lean proofs for some of them. OpenAI says the results are at different stages of checking and that some of the unchecked ones could have problems.

Until now, AI math claims came one at a time, and mathematicians could check each one. This is 722 papers in one drop, from a model nobody outside OpenAI can use, and OpenAI itself says some of them may be wrong. The checked-by-computer proofs are the solid part. The rest is a pile of homework for human mathematicians, who now have to sort real breakthroughs from mistakes. If even a handful hold up, this is a big moment for AI in science. But the burden of proof just got handed to the field, not carried by the company.

On Tuesday, 6 October 2026, OpenAI published “Sharing AI progress in mathematics.” The company says it is releasing a broad set of new math results from an internal model, and that the papers and the proof files are in a public repository on GitHub. An internal model is one OpenAI has not released. Nobody outside the company can run it. Those lines are OpenAI’s.

The repository is openai/math. Its readme says it holds math manuscripts and the files that go with the proofs, produced by that internal model. The catalog lists 722 manuscripts, grouped into 372 families. A family is a set of related papers. OpenAI’s examples are a main result, plus companion arguments, consequences, or a different proof of the same thing. 722 papers in 372 families means many results are published as more than one write-up. The overview that opens the catalog sorts the families into 17 subjects, including number theory, algebraic geometry, theoretical computer science, combinatorics, mathematical physics, and partial differential equations. Those counts are OpenAI’s.

OpenAI says it started giving its models open research problems after the models maxed out the math tests it already had. Maxed out means the scores had hit the ceiling, so those older tests no longer showed whether a newer model was any better. Over this evaluation it gave the model about 4,000 problems. On average, each result used about three hours of ChatGPT Pro thinking compute with that same model. ChatGPT Pro is OpenAI’s paid chat product. The three hours is a measure of computing time, not a person sitting with a clock. OpenAI then grouped what came out and kept the results it judged significant enough to publish. Those lines are in the repository readme.

Not everything is checked. OpenAI says the results are at different stages of verification. Not all of them have a Lean formalization. Lean is a programming language that lets a computer check every step of a proof. A passing check means the logic of that formalized claim holds. It does not, by itself, say the result is new or important. OpenAI says some of the results that have not been formalized “could have issues,” and that it will fix those quickly. The repository’s formalization catalog, the list of papers with a formally checked main result, currently lists 162 papers. 162 out of 722 is a bit under one in four. The other manuscripts do not yet have that computer check. Those lines and that count are OpenAI’s.

OpenAI also released shortened summaries of the model’s reasoning for 10 results. The readme’s ten include the irrationality exponent of pi, the Mahler conjectures, Kaplansky’s direct-finiteness conjecture in characteristic two, and the isomorphism of free group factors. A shortened summary is a cut-down account of the steps the model took. It is not the paper, and it is not a computer check of the proof. Those names are OpenAI’s.

OpenAI says most of the results came from the same fixed procedure, using that unreleased model. It names two exceptions. One is work on a zero-free region for the Riemann zeta function. That function is a formula tied to the primes, and a zero-free region is a part of the number plane where the formula is not zero. The other is a proof of the Hodge Conjecture for CM abelian varieties, a special case of a famous open problem in geometry, not the full conjecture. OpenAI says people edited the Riemann zeta write-up so it would be easier to read. Those lines are in the repository readme.

OpenAI says it is exploring community-hosted repositories for the material, places the math community would run rather than the company alone. It says it will keep every released version available, and that corrections will be logged as new versions instead of being written over the old ones. On the announcement, OpenAI also says it consulted the independent Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study, and used that group’s advice in deciding how to release these results. That advice is about the release. It is not a statement that the group has checked the 722 papers. Those lines are OpenAI’s.

Scientific American and Unite.AI both reported the release the same day. Scientific American’s headline is “OpenAI unleashes hundreds more math results upon a field already in shock.” Unite.AI’s headline is “OpenAI Releases 722 Math Manuscripts From an Unreleased AI Model.” Both describe the same public repository, the same unreleased model, and the same warning that some results are not yet checked. Those headlines are the outlets’.

The picture is page one of OpenAI’s Research Catalog, the overview that opens the repository. It lists 372 result families in 722 manuscripts. The contents on that page run from number theory through partial differential equations. The page sits on a dark teal backdrop. The frame does not print a calendar date. It is the catalog’s first page. It is not a photograph of a mathematician or a chalkboard.

In plain terms, OpenAI on Tuesday published 722 math manuscripts written by a model it has not released. They are grouped into 372 families across 17 subjects, drawn from about 4,000 problems the model was given. 162 papers so far have a main result a computer has checked. OpenAI says the rest are at different stages of checking, and that some of the unchecked ones could have problems. It also published shortened reasoning notes for 10 results, and it says old versions will stay online when it posts a correction.