OpenAI’s 722 Math Manuscripts: What the GitHub Release Contains

Quick answer: OpenAI has published a public GitHub collection containing 722 mathematical manuscripts organized into 372 families, along with source files, citation data, a formalization catalogue, Lean code for many results and selected reasoning summaries. The release is unusual because it combines large-scale AI-generated mathematical research with formal verification work, but it should not be read as “722 peer-reviewed breakthroughs.” Researchers still need to check novelty, assumptions, exposition and the correspondence between natural-language claims and any formal proof.

The official repository is github.com/openai/math. It is the best starting point because the repo itself explains how the collection is organized and which manuscripts have formalized counterparts.

OpenAI’s math release at a glance

What exactly did OpenAI publish?

The repository is more than a folder of PDFs. It includes an overview, a manuscript map that groups related results, a preprints area with paper files and citation information, a Lean library and formalization catalogue, and supporting material for reproducing or checking parts of the work.

The “372 families” figure is important. Multiple manuscripts can belong to the same broader line of investigation. That means counting every manuscript as an entirely independent discovery would overstate what the release represents. A family can contain related variants, refinements, companion results or papers that develop a shared mathematical direction.

How the repository is organized

A practical reading workflow looks like this:

1. Start with the repository overview to understand the project structure.
2. Use the manuscript map to find a research family rather than browsing hundreds of filenames blindly.
3. Open the corresponding preprint to read the theorem statement, assumptions, proof strategy and references.
4. Check whether the result appears in the formalization catalogue.
5. If Lean code exists, inspect the formal statement and dependencies rather than assuming the PDF and Lean theorem are automatically identical in scope.
6. Compare the bibliography and related literature before making claims about novelty.

What does a Lean proof actually prove?

Lean is an interactive theorem prover. When a theorem has been correctly formalized and Lean accepts its proof, that gives strong assurance that the formal statement follows from the encoded assumptions and definitions inside the formal system.

But formal verification does not automatically answer every scholarly question. A Lean proof does not by itself establish that:

• the formal theorem perfectly matches every sentence or scope condition in the natural-language manuscript;
• the result is new to mathematics;
• the result is important or useful;
• the manuscript’s motivation, attribution and literature review are complete;
• every intermediate informal claim in the paper has been formalized.

This distinction is one of the biggest gaps in short coverage of the release. “Machine-checked” is extremely meaningful, but it is not a synonym for “peer-reviewed, novel and publication-ready.”

Are all 722 manuscripts formally verified?

No. The repository provides formalization material for many results, but the official project structure does not present all 722 manuscripts as fully formalized end to end. Researchers should use the formalization catalogue to see which claims have Lean counterparts and what exactly has been checked.

A manuscript can also contain several propositions, corollaries, examples and contextual claims while only a central theorem has been formalized. The correct unit of verification is therefore the exact formal statement, not simply the paper title.

Does 722 manuscripts mean 722 breakthroughs?

No. The most defensible description is “722 manuscripts across 372 families.” Some may contain genuinely new results; some may refine known ideas; some may later turn out to overlap with existing literature; and some may be more valuable as formalization exercises or research leads than as major standalone discoveries.

That is normal for a research corpus of this size. The collection should be evaluated result by result, especially because public mathematical writing can move faster than conventional journal review.

How the manuscripts were generated

Reporting around the release says the manuscripts were produced with an unreleased OpenAI frontier system and substantial automated reasoning effort. Some coverage has cited an average on the order of a few hours of high-compute reasoning per result. That figure is better understood as a system-level compute description than as the wall-clock time a human mathematician would need to verify the work.

The more important research question is reproducibility: can other mathematicians understand the assumptions, rerun or inspect the formal material, identify prior art and validate the claimed contribution independently?

Why the 372-family structure matters

Grouping manuscripts into families helps reviewers avoid a common failure mode: evaluating related papers as though they were independent evidence. If ten manuscripts explore neighboring consequences of one construction, the family view makes that relationship visible.

It also helps specialists focus quickly. A researcher can identify a family close to their area, inspect the strongest or most central manuscript, then follow linked variants instead of reading hundreds of unrelated documents.

How a mathematician should audit one result

A rigorous audit can be broken into five checks:

1. Statement check: Is the theorem precise, and are all hypotheses explicit?
2. Literature check: Does an equivalent result already appear elsewhere under different terminology?
3. Proof check: Does the informal proof genuinely establish the stated result?
4. Formalization check: If Lean code exists, does the formal theorem match the intended natural-language claim?
5. Significance check: Even if correct and novel, does the result materially advance its field?

This is why the repository is useful even to skeptical readers: source files, formal artifacts and grouping metadata make systematic checking easier than it would be for a pile of anonymous PDFs.

How this differs from normal AI benchmark claims

Most frontier-model announcements are summarized through benchmark percentages. A mathematical corpus is different because individual outputs can be examined directly by domain experts and, in some cases, checked by a theorem prover.

That makes the OpenAI math release closer to a research artifact than a conventional model leaderboard. Readers comparing it with recent frontier-model announcements can also see AVARIXO’s Reflection Beam overview, Mistral Large 4 guide and GPT-6 Astra and GPT-6.1 Sol update.

Can the manuscripts be cited?

Yes, individual preprints can be cited in the normal scholarly sense if they provide stable authorship/title/version information, but researchers should cite the specific manuscript and version they actually used rather than citing “the 722 papers” as one undifferentiated result.

Because the collection is new and can evolve, record the commit, version or access date when reproducibility matters. If a manuscript later receives a revised version or formalization update, that distinction can become important.

What reviewers should be cautious about

External reporting has already highlighted debate over how quickly large volumes of AI-assisted mathematical work can be validated. That concern is reasonable. Formal methods can reduce one class of error, but they do not eliminate literature duplication, ambiguous informal interpretation or overstatement of significance.

The healthiest way to read the release is neither “AI solved mathematics” nor “AI-generated papers are worthless.” The repository is valuable precisely because it gives mathematicians concrete objects to inspect, criticize, formalize and compare.

FAQ

Did OpenAI publish 722 peer-reviewed papers?

No. OpenAI published 722 manuscripts/preprints in a public repository. Public release and formal verification are different from journal peer review.

What are the 372 families?

They are groupings of related manuscripts or research directions. A family may contain multiple connected results rather than representing multiple unrelated discoveries.

Are all 722 manuscripts proven in Lean?

No. The repository includes a Lean library and a formalization catalogue for many results, but readers should check the catalogue for the exact theorem they care about.

Does a Lean proof guarantee the manuscript is novel?

No. Lean checks formal logical correctness relative to encoded assumptions. Novelty requires comparison with prior mathematical literature.

Can anyone inspect the collection?

Yes. The repository is public on GitHub and includes the project structure, preprints and formalization resources needed for independent inspection.

Was the underlying model released?

The manuscripts have been associated with an unreleased frontier system. The public artifact is the research collection, not necessarily general access to the exact model that generated it.

Bottom line

OpenAI’s 722-manuscript release is significant because it exposes a large body of AI-assisted mathematical work to direct scrutiny rather than asking the public to trust a single benchmark number. The right way to evaluate it is family by family and theorem by theorem: read the manuscript, inspect the formal statement where available, check prior art and separate machine-verified correctness from novelty and scientific significance.