OpenAI Math: 722 Manuscripts Released for Review

OpenAI math now includes 722 mathematical manuscripts generated by an unreleased internal AI model, giving mathematicians access to a huge collection of proposed solutions and proofs as the company faces growing scrutiny over how AI-generated discoveries should be verified and credited. The manuscripts are organized into 372 families of related results and are available through a public GitHub repository.

The release is significant not simply because of its scale, but because it comes after weeks of criticism from mathematicians over how frontier AI companies announce mathematical breakthroughs. OpenAI says it consulted an independent Advisory Group on Mathematics and Artificial Intelligence when deciding how to release the work, while the advisory group has stressed that publication is only the beginning of the process and does not constitute an endorsement of the results or the way they were produced.

Source and verification note (7 October 2026): OpenAI’s 6 October announcement and the public repository establish the 722-manuscript, 372-family counts. Independent original examinations by The Information, CellCog and Kingy AI corroborate the release and examine its limits. The independent mathematics advisory group does not certify the papers. Manuscript publication is not proof of 722 verified breakthroughs.

Key takeaways

• OpenAI has released 722 mathematical manuscripts grouped into 372 result families.
• The work was generated primarily by an unreleased internal OpenAI model.
• OpenAI says approximately 4,000 mathematical problems were presented to the model during the evaluation.
• The average result used compute equivalent to roughly three hours of ChatGPT Pro thinking.
• Many manuscripts have accompanying Lean formalizations, allowing portions of the proofs to be checked computationally.
• OpenAI has also released 10 abbreviated reasoning summaries and additional information about its evaluation process.
• The repository explicitly warns that some unformalized results could contain errors.
• The release follows criticism from mathematicians over verification, attribution, transparency and the commercialization of mathematical discoveries.
• The independent advisory group says the mathematical community still needs to determine how significant and correct the results actually are.

OpenAI math release: 722 manuscripts from an internal model

The OpenAI math release on October 6 represents one of the largest public disclosures of AI-generated mathematical research by a major AI company.

The company’s public repository contains 722 manuscripts arranged into 372 families. A family can include a primary result alongside companion arguments, consequences or alternative proofs. The collection covers multiple areas of mathematics rather than focusing on a single benchmark or mathematical discipline.

OpenAI describes the material as mathematical manuscripts and supporting proof artifacts produced by an internal model. The model itself has not been released publicly.

The company says most of the results came from a common evaluation procedure. Its internal model was presented with approximately 4,000 open research problems, and OpenAI subsequently selected outputs that met what it considered an appropriate level of significance.

That distinction is important. The 722 manuscripts are not 722 unrelated problems that were all solved from scratch. The company says the manuscripts were aggregated into result families, meaning several papers can relate to the same underlying mathematical development.

OpenAI also says some outputs build upon earlier results produced by its models.

The AI model behind the research remains private

One of the biggest unanswered questions is the identity and capability of the model that generated the work.

OpenAI has not released the model itself. The company describes it as an internal frontier model and says it is working toward a responsible release.

That creates a fundamental difference between this collection and an ordinary academic research archive.

A mathematician reading one of the papers can inspect the proposed proof, reproduce calculations where possible and attempt independent verification. However, researchers cannot independently reproduce the entire discovery process by running the same model because the underlying system is not publicly available.

OpenAI has attempted to compensate for some of that limitation by publishing additional information about the process.

The company says the average result required computing equivalent to approximately three hours of ChatGPT Pro thinking. It has also disclosed that the model was given around 4,000 problems during the evaluation.

These figures provide a rough indication of the computational resources involved, but they do not make the underlying AI research process fully reproducible.

Lean formalization gives mathematicians another way to check results

A particularly important component of the release is the use of Lean.

Lean is a formal programming language and proof system that can allow mathematical arguments to be checked by computer. Instead of relying solely on a human reader to examine whether every logical step follows correctly, a formalized proof can be processed by a proof assistant.

OpenAI says many of the newly released results have Lean formalizations and that it will add more as they become available.

However, formalization does not mean every paper in the collection has already been independently validated.

OpenAI’s repository explicitly says the collection contains results at different stages of verification. Not every manuscript has a corresponding Lean formalization, and the company acknowledges that some unformalized results could contain issues.

That caveat is central to understanding the release.

The appropriate interpretation is therefore not that OpenAI has conclusively established 722 new mathematical breakthroughs. Instead, the company has made 722 manuscripts publicly available for examination, with varying levels of machine-assisted verification.

Why the release matters for mathematical research

The scale of the collection could change how mathematicians interact with AI-generated research.

Historically, a mathematical result normally moves through a relatively slow process. Researchers develop a conjecture or proof, write a paper, discuss it with colleagues, submit it for publication and undergo some form of peer review or community scrutiny.

AI systems are changing the speed at which candidate results can be generated.

If a frontier model can produce hundreds of potentially meaningful mathematical arguments, the bottleneck could shift from generating proofs to checking, understanding and contextualizing them.

That creates an unusual problem for mathematicians.

A result can be technically correct but still difficult to understand. A proof may establish a statement without making clear why the result matters, how it connects to existing theory or what new ideas it introduces.

This distinction between proving something and understanding something has become a central issue in the debate over AI mathematics.

The advisory group has argued that the long-term future of mathematics cannot simply become a process in which researchers spend their time digesting results generated by AI laboratories. Human mathematicians must continue to formulate questions, develop approaches and explore areas that AI systems have not selected as benchmarks.

The credit problem is bigger than simply naming an AI model

The controversy surrounding OpenAI’s mathematics work is partly about attribution.

When an AI system produces a mathematical result, determining who deserves credit can become complicated.

The model may have been trained on a large body of human-created mathematical knowledge. Researchers may have contributed ideas or prompts. Engineers developed the model. Mathematicians may have previously explored related problems. And an AI-generated proof may combine techniques that already exist in the literature in a new way.

That makes the traditional concept of authorship difficult to apply.

The issue became particularly prominent after OpenAI announced its claimed solution to the Navier–Stokes Millennium Prize problem in September.

Mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge had separately been working on related mathematical questions with the assistance of AI tools. Questions were subsequently raised publicly about whether prior human work could have influenced OpenAI’s result.

OpenAI has denied that Buckmaster’s prompts influenced its Navier–Stokes result and said it investigated the issue. The company also said its proof differed significantly from the work by Buckmaster and Alpöge.

The dispute illustrates why AI-generated scientific research requires detailed records of provenance, interactions and prior work.

Nature subsequently argued that AI companies need to work with the research community to develop reliable ways to protect attribution and research integrity.

OpenAI created an independent mathematics advisory group

The latest release did not happen in isolation.

In September, OpenAI announced an independent Advisory Group on Mathematics and Artificial Intelligence associated with the Institute for Advanced Study.

The group includes prominent mathematicians such as Timothy Gowers, Martin Hairer, Ravi Vakil and Edward Witten.

The group was established after discussions between OpenAI and mathematicians about how AI companies should interact with mathematical research.

Importantly, the advisory group says it is independent of OpenAI. Its members are not paid by AI companies for the advisory work, and the group does not have decision-making authority over OpenAI.

The group’s recommendations were shaped by more than 600 survey responses, according to its website.

OpenAI says it used the group’s advice when preparing the 722-manuscript release.

But the advisory group has been careful not to endorse the collection itself.

In its statement following the release, the group said its advisory role should not be interpreted as a judgment on the impact of the results or an endorsement of the process through which OpenAI obtained them.

Instead, it described the publication as a first step toward the much larger task of human understanding and incorporation of the results into mathematical knowledge.

OpenAI is publishing more information about how the results were created

The company has attempted to make the release more transparent than a conventional AI product announcement.

Alongside the manuscripts, OpenAI has published information about its evaluation process, supporting proof artifacts and selected reasoning summaries.

The repository contains 10 abbreviated summaries of the model’s reasoning for selected results.

The company has also established protocols for revisions and citations. Earlier versions are intended to remain accessible so researchers can track how individual manuscripts change over time.

This is particularly important for AI-generated research because errors can be discovered after publication.

Traditional academic papers can also contain errors, of course. The difference is that the speed and volume of AI-generated research could make the correction process much more important.

If hundreds or thousands of mathematical manuscripts are generated rapidly, researchers need a reliable way to identify which results have been checked, which have been challenged and which have been corrected.

OpenAI’s versioning system is an attempt to provide that infrastructure.

The repository is not the same as peer-reviewed publication

Another important distinction is between making research publicly accessible and having it accepted by the mathematical community.

The 722 manuscripts are available online, but their publication on GitHub does not mean they have gone through conventional peer review.

This is one of the concerns raised by mathematicians.

AI-generated mathematical work can be difficult to evaluate because researchers may need substantial time to understand an argument before they can determine whether it is genuinely new, correct and important.

The sheer volume of OpenAI’s release could therefore create a paradox.

Making the papers public makes them easier to inspect, but publishing hundreds of results simultaneously may also place a substantial verification burden on the mathematical community.

The advisory group itself has said that the community still needs to undertake that assessment.

What the release says about the future of AI research

The bigger significance of the 722-paper release may be less about any individual theorem and more about the changing economics of scientific discovery.

Frontier AI companies increasingly have the computing power to run models against difficult research problems for extended periods.

OpenAI says its evaluation involved around 4,000 problems and thousands of hours of equivalent reasoning compute when the results are considered collectively.

That creates a new research model: an AI laboratory can search through enormous numbers of mathematical possibilities and then publish the most promising outputs.

Human researchers could increasingly become validators, interpreters and collaborators rather than the sole generators of candidate solutions.

That could be highly productive if the systems produce reliable insights.

But it could also create problems if companies control the most capable research systems while the wider academic community is left with the responsibility of checking their outputs.

This is why questions around access, attribution and transparency are becoming as important as raw model performance.

What happens next

The immediate next step is independent scrutiny.

Mathematicians will need to examine individual manuscripts, compare them with existing literature and determine which arguments are correct and genuinely novel.

Lean formalizations can accelerate this process for the results that have been formalized, but human mathematical judgment remains necessary to establish significance and context.

OpenAI has also said it plans to fund workshops, conferences and special programs focused on understanding major results produced by AI.

The company says it will continue improving its standards for future disclosures and is exploring community-hosted alternatives to its GitHub repository.

The advisory group, meanwhile, has indicated that its discussions with OpenAI will continue but has emphasized that the mathematical community must ultimately decide how the results should be incorporated into the field.

For readers tracking AI research, see our coverage of OpenAI’s Decisions API and the separate Codex Auto-Review update. Those commercial product launches are distinct from this unreleased mathematics model.

The Bigger Picture

OpenAI’s 722-manuscript release marks a shift from AI demonstrating mathematical ability through isolated benchmark results to AI producing a body of research large enough to resemble a small research program.

That changes the central question. Instead of asking only whether AI can solve difficult mathematics, researchers now have to ask how society should verify, publish, attribute and build upon mathematics produced by AI systems.

The answer will likely involve a combination of machine-checkable proofs, conventional peer review, better provenance records and greater access to the underlying research process. If those mechanisms develop alongside AI capability, models could become powerful research partners. If they do not, the industry risks producing more mathematical results than the scientific community can confidently evaluate.

Frequently Asked Questions

Did OpenAI really publish 722 math papers?

Yes. OpenAI’s public repository contains 722 mathematical manuscripts organized into 372 related result families. The company describes them as manuscripts and supporting proof artifacts generated primarily by an unreleased internal model.

Are all 722 mathematical results proven to be correct?

No. OpenAI explicitly says the manuscripts are at different stages of verification and that not all have Lean formalizations. The company also warns that some unformalized results could contain issues.

Did OpenAI release the AI model that created the papers?

No. The model remains an internal system. OpenAI says it is working toward responsibly releasing the model but has not announced a public release of it alongside the manuscripts.

Why is credit such a major issue?

AI-generated mathematical discoveries can draw on existing human knowledge while also involving model developers, researchers and users. Determining whether a result is genuinely new, what previous work influenced it and who should receive academic credit is considerably harder when the discovery process is partly opaque.

Looking Ahead

The 722 manuscripts will now enter a much slower phase: human scrutiny. The most important results will need to survive independent mathematical examination before they become accepted contributions to the field. Some may prove important, some may require corrections and others may turn out to be less significant than the initial AI-generated claims suggest.

For OpenAI, the episode is also a test of whether frontier AI research can coexist with established scientific norms. Publishing the work openly, preserving revisions, providing machine-checkable proof artifacts and engaging an independent advisory group are steps toward greater transparency, but the ultimate measure will be whether mathematicians can independently verify the work and receive appropriate credit for the ideas that make it possible.