OpenAI Math Repo: 722 AI-Produced Manuscripts on GitHub

Introduction

On October 6, 2026, OpenAI published a GitHub repository, openai/math, containing what it calls "a broad range of new mathematical results produced by an internal frontier model." The README describes the contents as mathematical manuscripts and supporting proof artifacts, released under the Apache-2.0 license. The model is not named and has not been released; OpenAI says it is working to "responsibly release the model that produced these results."

The release follows OpenAI's earlier announcement of a Navier-Stokes result, which this site covered separately. This article focuses on the batch release itself and on how OpenAI chose to disclose it. Several parts of the release matter for readers: the repository's size, the verification status of the results, and the process OpenAI says it followed after consulting outside mathematicians.

Feature Overview

Scale and organization. According to the README, the repository holds 722 manuscripts organized into 372 families. A family groups a principal result with companion arguments, consequences or alternative proofs, and the families are classified by discipline. OpenAI says it posed the model approximately 4,000 problems; aggregating the output and "requiring an appropriate level of significance" produced the catalog. The selection is therefore OpenAI's own.

Verification status. The README says the results are "at different stages of verification. Not all have accompanying Lean formalizations," and warns that "some of the unformalized results could have issues." Lean is a proof assistant that mechanically checks a formal proof. OpenAI's blog says many proofs have Lean formalizations and that more are coming. The 722 manuscripts should not be read as verified or peer-reviewed.

Reasoning summaries and compute. OpenAI is publishing 10 summaries of the model's reasoning, compute estimates expressed in ChatGPT Pro usage, and statistics on the number of attempted problems. It states that "the average result used the equivalent compute of roughly three hours of ChatGPT Pro thinking." The reasoning-summary families named in the README include the irrationality exponent of pi, the symmetric and general Mahler conjectures, Kaplansky's direct-finiteness conjecture in characteristic two, the Mézard–Parisi formula for diluted spin glasses, and the isomorphism of free group factors.

Exceptions to the fixed procedure. The README names two lines of work that departed from the standard procedure. One is a zero-free region for the Riemann zeta function. The other is a proof of the Hodge Conjecture for CM abelian varieties. These are the stated scopes: a zero-free region, not the Riemann Hypothesis, and a CM abelian varieties case, not the general Hodge Conjecture. The README adds that the Re(s) > 11/12 zero-free region writeup "was human edited for readability."

Versioning. The release history is preserved, and corrections are recorded as new versions. OpenAI says the repository includes protocols for paper revisions and citations.

Usability Analysis

For mathematicians, the practical value lies in the artifacts. A manuscript with an accompanying Lean formalization can be checked mechanically, while an unformalized one needs human review. The classification by discipline and the family structure help a specialist locate relevant results, and the reasoning summaries offer a view into how the model approached a handful of problems.

The disclosure practice is also notable. OpenAI says it consulted the independent Advisory Group on Mathematics and Artificial Intelligence (AGMAI) at the Institute for Advanced Study and drew on its public recommendations. The Verge reports that AGMAI's late-September recommendations urged labs to release results promptly and through established academic channels where possible, to disclose the model name, prompts and compute costs, and to "refrain from treating the release of mathematical results as marketing vehicles to promote their models." OpenAI has published compute estimates and attempt statistics, but it has not named the model.

Pros and Cons

The strengths are openness of the artifacts, an explicit verification caveat, preserved version history and the Lean proofs. The weaknesses are the absence of journal review, an unnamed model that outsiders cannot test, and a catalog selected by the vendor from roughly 4,000 attempts. The Verge frames the release as extending breakthroughs that "both impressed and unsettled parts of the mathematical community while raising questions about research ethics and academic conduct."

Outlook

OpenAI says it is still exploring community-hosted alternatives to GitHub that meet the committee's guidelines, and that it will fund workshops, conferences and special programs on understanding AI-produced results. The Verge reports that AGMAI says the release includes solutions to "hundreds" of open questions, and that OpenAI said in September its model had "resolved more than 100 long-standing open problems across most areas of mathematics." Those are statements from AGMAI and OpenAI, not independent counts. The next signals to watch are how many manuscripts receive Lean formalizations, whether outside mathematicians confirm or refute specific results, and whether the model becomes available for others to test.

Conclusion

The repository is a substantial, openly licensed dataset of AI-produced mathematics, released with more caveats and process than a typical model announcement. It is most useful to researchers who can evaluate proofs and to anyone studying how AI labs should disclose scientific claims. Readers should treat each result as a claim awaiting verification, particularly the unformalized ones, and should keep the exact scope of the two exceptional results in mind.

Editor's Verdict

OpenAI Math Repo: 722 AI-Produced Manuscripts on GitHub earns a solid recommendation within the GPT space.

The strongest case for paying attention: an Apache-2.0 license lets researchers reuse and study the manuscripts and proof artifacts. That alone raises the bar for what readers should expect in this space. Reinforcing that, the README states plainly that not all results have Lean formalizations and that some could have issues — practical value rather than just headline appeal. The broader signal worth registering is straightforward: Lean formalization turns a claimed proof into something mechanically checkable, which is why its coverage matters more than the manuscript count. On the other side of the ledger, one constraint is real rather than a marketing footnote: the results bypass journal peer review for now, and not every manuscript has a Lean formalization yet. It should factor into any serious decision. Layered on top of that, the catalog was selected by OpenAI from roughly 4,000 posed problems, so success rates and selection criteria rest on the vendor's account — which narrows the set of teams for whom this is an obvious yes.

For research mathematicians, Lean users, and anyone tracking how AI labs disclose scientific claims, this is a serious evaluation candidate, not just a curiosity to bookmark. For everyone else, the safer posture is to monitor coverage and revisit once the use cases that matter to your team are demonstrated in the wild.