OpenAI Releases 722 Math Papers From an Unreleased AI Model
OpenAI has released an extraordinary collection of mathematical research: 722 manuscripts organized into 372 distinct result clusters, all generated by an internal model that remains unavailable to the public.
The catalog spans several major areas of contemporary mathematics, including the quasi-Riemann hypothesis, the Unique Games Conjecture, special cases of the Hodge conjecture, and the free group factor problem. Some results also include Lean formalization materials intended to make key arguments machine-checkable.
The scale of the release is significant, but the number 722 should not be interpreted as 722 independently solved hard problems. A single result cluster can contain a primary theorem, supporting arguments, corollaries, alternative proofs, and related manuscripts.
OpenAI says the work originated from roughly 4,000 mathematical problems. Results were subsequently grouped and filtered according to their significance, producing the public catalog. On average, each result reportedly consumed computational effort equivalent to around three hours of ChatGPT Pro thinking.
The manuscripts are public, but the model that generated them is not. That distinction is important: mathematicians can now inspect the proposed results, but they cannot generally reproduce the same research process or directly interrogate the underlying system.
The full collection is available on GitHub:
https://github.com/openai/math🧮 A Massive Catalog of AI-Generated Mathematics #
OpenAI describes the release as part of a broader shift toward AI-assisted mathematical discovery.
The company has argued that conventional mathematical benchmarks are approaching saturation, making long-standing open research problems a more meaningful test of frontier reasoning systems. Rather than evaluating a model solely on problems with known answers, the new approach asks whether AI can generate genuinely useful mathematical results.
The public catalog therefore represents something different from a conventional benchmark leaderboard.
It is a research corpus that researchers can inspect, challenge, formalize, extend, or potentially refute.
Sam Altman characterized the development as entering a “new era of discovery.” The more consequential question, however, is not simply how many papers an AI system can produce, but whether mathematicians can verify the arguments and extract reusable ideas from them.
That distinction becomes especially important for the most ambitious results in the collection.
🔢 The Quasi-Riemann Hypothesis Result #
Among the most attention-grabbing entries is a manuscript concerning a quasi-Riemann hypothesis.
Mathematician Alex Kontorovich of Rutgers University reacted to the result on X with evident surprise, highlighting the significance such a result would have if produced and verified by a human mathematician.
To understand why, consider the underlying problem.
The Riemann zeta function is deeply connected to the distribution of prime numbers. Its non-trivial zeros occur within the critical strip, and the classical Riemann hypothesis asserts that every non-trivial zero has real part exactly equal to 1/2.
The new manuscript does not prove the Riemann hypothesis.
Instead, it proposes a substantially narrower zero-free region: there are no zeros in the region where the real part is greater than 7/8.
That is why the “quasi” qualifier matters.
A fixed-width zero-free region #
Known zero-free results for the zeta function become increasingly subtle as the imaginary component grows. Kontorovich noted that even expanding known zero-free regions around the right side of the critical strip could constitute meaningful progress.
The manuscript instead proposes a fixed-width region extending indefinitely within the critical strip.
In simplified terms, the claim is that the entire region
\[ \operatorname{Re}(s) > \frac{7}{8} \]contains no non-trivial zeros.
If correct, this would be a substantial mathematical result. It nevertheless falls well short of proving the full Riemann hypothesis, which requires establishing that all non-trivial zeros lie on the line
\[ \operatorname{Re}(s) = \frac{1}{2}. \]The repository also includes Lean formalization materials associated with the result.
That makes the manuscript particularly interesting, because researchers can inspect not only the mathematical argument but also portions of its machine-checkable representation.
🧩 Unique Games Conjecture and Approximation Hardness #
Another major claim concerns the Unique Games Conjecture (UGC) and its implications for approximation algorithms.
Many computational optimization problems are difficult to solve exactly at scale. Instead, algorithms often seek solutions that are guaranteed to be within a certain factor of the optimum.
Max-Cut is a familiar example. Given a graph, the objective is to divide its vertices into two sets while maximizing the number of edges crossing between them.
The relevant papers claim a proof of the Unique Games Conjecture and derive approximation hardness consequences for problems such as Max-Cut.
If these arguments survive verification, they would establish important limits on what efficient algorithms can guarantee for general instances, particularly under the additional assumption that P ≠ NP.
These are worst-case computational guarantees, however. They do not imply that every practical dataset is difficult or that better performance cannot be achieved for restricted graph families, structured inputs, or application-specific instances.
That distinction is critical when translating theoretical approximation hardness into practical algorithm design.
🧬 A Partial Result on the Hodge Conjecture #
The catalog also includes work related to the Hodge conjecture, another of the famous Millennium Prize Problems.
At a high level, the Hodge conjecture asks whether certain topological structures in complex algebraic varieties can be represented through algebraic geometric objects defined by polynomial equations.
The newly reported result focuses on the rational Hodge conjecture for CM abelian varieties.
This is a specific but mathematically significant class of objects. It should not be confused with a complete solution to the Hodge conjecture.
In other words, even a correct proof of this result would represent progress on an important special case rather than the resolution of the full Millennium Prize Problem.
The current manuscript also does not list accompanying Lean formalization materials in the same way as some of the other highlighted results.
🔬 The Free Group Factor Problem #
Another striking result concerns the free group factor problem, which lies within operator algebras and von Neumann algebra theory.
The underlying question can be framed around whether certain operator-algebraic constructions arising from free groups can be equivalent in the relevant structural sense.
The manuscript reportedly answers this question affirmatively and extends the result to interpolated free group factors with parameters greater than 1, including infinite-parameter cases.
Lean formalization materials are also available for this result.
One important technical distinction is that equivalence of the constructed operator algebras does not mean that the underlying free groups themselves become identical groups. The result concerns the associated operator-algebraic structures.
✅ Lean Formalization Helps, but It Is Not a Universal Stamp of Approval #
One of the most consequential aspects of the release is the presence of Lean formalization.
According to the public catalog, 162 manuscripts have main conclusions accompanied by Lean formalization materials.
Formalization translates mathematical definitions, propositions, and proofs into a machine-checkable language. Lean can then verify that the encoded deductions follow according to its formal logic and that required proof obligations have been discharged.
This is considerably stronger than simply asking another language model whether a proof “looks correct.”
But formal verification has an important boundary: the formalized statement must correspond to the mathematical claim researchers actually intended to prove.
Formalization cannot fix an incorrect premise #
Suppose a natural-language theorem says that a property holds for every object in a given class, while the formalized version silently adds an extra assumption.
Lean may correctly verify the formalized theorem without establishing the broader claim.
The machine can verify the proof it receives. It cannot independently determine whether the formal statement faithfully represents the original mathematical problem.
This is especially relevant to AI-generated mathematics, where subtle changes in definitions, quantifiers, assumptions, and scope can materially alter a theorem.
The quasi-Riemann manuscript provides a useful example. Its formalization covers important portions of the argument, but the notes explicitly indicate that subsequent applications of the paper fall outside the formalization’s scope.
Consequently, “has Lean formalization” should not be interpreted as “every statement in the paper has been formally verified.”
OpenAI also acknowledges that some results without formalization may contain errors and are subject to revision.
🧠 Verification Is Only the Beginning #
Even when a mathematical deduction passes formal verification, researchers still need to understand why the method works.
A proof assistant can establish that a formal proof follows from its definitions and axioms. It does not automatically explain which conceptual insight produced the argument, why the approach is useful, or where the same technique might apply elsewhere.
That distinction matters enormously for mathematical research.
A result that solves one isolated problem may be less valuable than a new technique that allows mathematicians to solve an entire class of problems.
This is one reason the current release is likely to become a substantial research project in its own right.
👨🔬 OpenAI and Mathematicians Have Been Debating This for Months #
The release follows months of increasingly intense interaction between OpenAI and the mathematics community.
In August, OpenAI gathered roughly 40 mathematicians and suggested that its internal system had solved hundreds of problems. Participants reportedly requested papers that could be independently evaluated.
The discussions also produced disagreement over whether OpenAI had committed to releasing the results in a particular form.
On September 8, OpenAI announced what it described as a solution to the Navier–Stokes Millennium Prize Problem, accompanied by a paper and Lean formal proof.
Three days later, 25 Fields Medalists, with Terence Tao among the initial signatories, published an open letter criticizing what they viewed as a mismatch between AI development priorities and the broader goals of mathematical research.
The letter emphasized that mathematical progress is not solely about obtaining answers to open problems. Researchers also value new concepts, methods, citations, understanding, and the development of future mathematical talent.
The concern is therefore not simply that AI may solve problems too quickly. It is that the research ecosystem may struggle to absorb a rapidly increasing volume of machine-generated results.
📚 The Mathematicians’ Response Is Not Uniform #
The mathematics community has not responded with a single position.
Timothy Gowers, for example, suggested in a September 17 blog post that even an influx of thousands of major mathematical results could potentially be distributed among researchers by specialty and gradually analyzed.
His greater concern was the potential effect on the academic environment that currently supports mathematical research and the human effort required to understand new results.
An independent mathematics and AI advisory group, AGMAI, subsequently emerged amid these discussions.
OpenAI announced on September 21 that a new internal model, trained beginning August 28, had solved more than 100 long-standing open problems. At that stage, however, the company did not provide an itemized catalog of those results or their associated proofs.
On September 29, AGMAI published recommendations that included improving citations, writing papers for peer mathematicians, submitting work to independent preprint archives, and funding researchers who can digest and evaluate AI-generated discoveries.
The group also argued against relying exclusively on proprietary models that remain inaccessible to the broader mathematical community for frontier-level mathematical research.
🌐 722 Papers Change the Verification Problem #
OpenAI’s latest release represents a substantial change in that situation.
Instead of announcing a handful of headline results, the company has now made 722 manuscripts publicly accessible and organized them into 372 result clusters.
OpenAI has also established revision and citation standards, released 10 reasoning summaries, and committed funding toward related workshops. The company says it is continuing to explore community hosting arrangements consistent with recommendations from the mathematics community.
AGMAI clarified on the same day that its advice should not be interpreted as an evaluation of the mathematical impact of the results or an endorsement of the way they were generated.
That distinction matters.
Making the manuscripts public creates an opportunity for independent scrutiny. It does not establish that the results are correct.
⚠️ The Unreleased Model Is the Missing Piece #
There is still a major asymmetry in the current situation.
Researchers can download and inspect the generated manuscripts, but the underlying model remains unavailable to the public.
That means mathematicians can evaluate the outputs without generally having access to the system that produced them.
They cannot simply submit a new conjecture, ask the model to explain an unexpected step, request an alternative proof, or investigate a related problem using the same system.
This limits reproducibility in a way that differs from conventional mathematical research.
A published paper normally provides enough information for another researcher to reproduce the argument independently. An AI-generated research corpus additionally raises the question of whether researchers should have access to the discovery system itself.
🔭 The Real Test Starts After the Release #
The 722 manuscripts may ultimately prove more important as a research dataset than as a collection of headlines.
Mathematicians now have to determine which results are correct, which require revisions, which contain genuinely new techniques, and which can be generalized.
The Lean formalizations provide an especially valuable verification layer, but they cover only part of the catalog and do not eliminate the need for human mathematical judgment.
The most important development may therefore come later, when researchers gain broader access to the underlying AI system and can use it to investigate their own questions.
That would change the relationship between mathematicians and AI from one-way consumption of machine-generated answers to interactive mathematical research.
Without that shift, there is a potential role reversal: instead of AI becoming a research assistant for mathematicians, mathematicians could find themselves becoming verification assistants for AI-generated research.
The publication of 722 manuscripts is therefore less a final demonstration than the beginning of a much larger experiment.
The difficult question is no longer simply whether an AI system can produce advanced mathematics.
It is whether human mathematicians and AI systems can build a research process in which discovery, verification, understanding, and new questions all move forward together.
