OpenAI’s 722 AI-Generated Math Results: What We Know

OpenAI’s 722 AI-Generated Math Results: What We Know

OpenAI has released a remarkable catalog of mathematical research: 722 manuscripts organized into 372 result families, reportedly generated by an unreleased internal frontier model.

The catalog spans 17 areas of mathematics and theoretical computer science, including number theory, combinatorics, algebraic geometry, probability, differential geometry, mathematical physics, operator algebras, and mathematical logic.

More strikingly, several entries claim progress on problems that have remained open for decades, including the quasi-Riemann hypothesis, 4D Kakeya conjecture, Unique Games Conjecture, isomorphism problem for free group factors, Hilbert’s Tenth Problem over the rational numbers, Hodge conjecture for CM abelian varieties, and Mahler conjecture.

The scale is almost as significant as the individual claims. According to OpenAI, the catalog emerged from an evaluation in which the model was given approximately 4,000 research problems. After filtering and grouping related outputs, 372 sufficiently important result families remained, producing 722 manuscripts.

Sam Altman described the moment as the beginning of a new era of discovery.

But the most important question may not be whether an AI can produce hundreds of mathematical results.

It may be whether mathematicians can understand, verify, formalize, and build upon them fast enough.

🚨 722 Manuscripts From a Single Mathematical Evaluation

#

OpenAI says the release originated as an evaluation of the mathematical capabilities of an internal frontier model.

Existing mathematical benchmarks had largely saturated, so the evaluation was expanded toward genuine open research questions. Some results also built upon earlier outputs generated by the same model.

The evaluation reportedly involved roughly 4,000 research problems.

After clustering related outputs into result families and filtering them according to an “important enough” threshold, the final catalog contained:

• 372 result families
• 722 manuscripts
• 17 mathematical and theoretical-computer-science fields
• Approximately 4,000 attempted research problems
• An average compute budget equivalent to roughly three hours of ChatGPT Pro thinking per result

A result family generally contains a main theorem or claim together with supporting arguments, corollaries, variants, or alternative proofs.

The distribution is broad. The largest categories include:

The manuscript timestamps provide another striking clue.

Most folders are dated between September 23 and October 5, with September 23 and 24 alone accounting for nearly 200 papers each. That suggests the overwhelming majority of this catalog was generated in a highly concentrated period of less than two weeks.

This is not simply a collection of isolated AI answers. It looks much more like a large-scale automated research campaign.

🔬 What Was Actually Released?

#

OpenAI did not release only manuscript PDFs.

The repository also contains several forms of supporting material intended to make the results easier to evaluate.

Three particularly important artifact types are:

1. Lean formalizations
2. Abridged reasoning summaries
3. Compute and evaluation statistics

The Lean component is especially significant.

Based on the repository catalog, 235 of the 372 result families include a Lean formalization overview, representing roughly 63% of the families.

Formal verification can dramatically reduce the cost of checking mathematical arguments. Instead of requiring a human expert to manually inspect every inference, a formally encoded proof can be checked by a proof assistant.

However, formalization has an important limitation: it verifies the formal statement that was encoded. It does not necessarily establish that the formal statement captures the mathematical claim researchers originally intended.

OpenAI itself acknowledges that not every manuscript has been formalized and that unformalized results may contain errors.

That distinction becomes critical when evaluating the most ambitious claims in the catalog.

🧮 The Quasi-Riemann Hypothesis

#

The first result to attract widespread attention was item #003, concerning the quasi-Riemann hypothesis.

The classical Riemann hypothesis states that every non-trivial zero of the Riemann zeta function has real part

The full conjecture remains unresolved.

The quasi-Riemann problem asks for something weaker but still extraordinarily difficult: establish a constant \(\theta < 1\) such that the zeta function has no zeros in the half-plane

A fixed zero-free half-plane of this kind would represent a major improvement over known results.

The OpenAI result claims

uniformly for the Riemann zeta function and all Dirichlet \(L\)-functions.

The associated family also reports a uniform exclusion result for Landau–Siegel zeros.

Most importantly, this particular result already has a Lean formalization.

That does not mean the mathematical community should immediately treat the underlying research program as settled. It does, however, make this result substantially easier to audit than a purely informal proof.

📐 The 4D Kakeya Problem

#

Another major entry is item #074, concerning the Kakeya problem.

The Kakeya conjecture has been one of the central problems connecting harmonic analysis, geometric measure theory, and combinatorics.

The three-dimensional case recently became particularly prominent after Hong Wang and Joshua Zahl’s work on the 3D Kakeya set conjecture.

The OpenAI catalog now claims a result that simultaneously resolves:

• the 3D Kakeya maximal function conjecture, and
• the 4D Kakeya Hausdorff dimension conjecture.

The second claim is especially significant because higher-dimensional Kakeya problems have remained important targets following progress in lower dimensions.

Unlike the quasi-Riemann result, however, the 4D Kakeya claim is among the major results for which complete formal verification is not yet available.

That makes independent expert verification particularly important.

💻 Unique Games and Theoretical Computer Science

#

The catalog is not dominated by pure mathematics.

Item #102 reportedly claims a proof of the Unique Games Conjecture, originally proposed by Subhash Khot in 2002.

The result is particularly consequential because the Unique Games Conjecture has deep implications for the hardness of approximation.

According to the catalog, the claimed proof is then used to derive optimal approximation-hardness bounds for problems including:

• Max-Cut
• Vertex Cover
• Other classical optimization problems

If correct, this would be a major theoretical computer science result rather than merely another isolated theorem.

🧬 Hodge, BSD, and Hilbert’s Tenth Problem

#

Several other entries target famous problems across algebraic geometry and number theory.

Rational Hodge conjecture

#

Item #032 claims a proof of the rational Hodge conjecture for all complex CM abelian varieties.

The Hodge conjecture is one of the Clay Mathematics Institute’s Millennium Prize Problems, and even specialized cases can involve extremely deep interactions between algebraic geometry, topology, and number theory.

Birch and Swinnerton-Dyer

#

Items #002 and #006, taken together, reportedly establish a full Birch and Swinnerton-Dyer formula on a quadratic-twist family of density one for every rational elliptic curve.

The Birch and Swinnerton-Dyer conjecture is another Millennium Prize Problem.

Hilbert’s Tenth Problem over \(\mathbb{Q}\)

#

Item #004 reportedly gives a negative answer to Hilbert’s Tenth Problem over the field of rational numbers.

The original Hilbert’s Tenth Problem asks whether there exists an algorithm that determines whether a polynomial equation with integer coefficients has an integer solution.

The analogous question over the rational numbers has remained open despite decades of work.

A correct resolution would therefore be a landmark result in mathematical logic and number theory.

🧠 Free Group Factors and Other Major Claims

#

The catalog contains similarly ambitious claims outside number theory and algebraic geometry.

Item #287 reportedly resolves the isomorphism problem for free group factors, concluding that free group factors of different ranks are mutually isomorphic.

The result is accompanied by Lean formalization.

Other reported results include claims that:

• the irrationality measure of \(\pi\) is exactly 2;
• Catalan’s constant is irrational;
• the Mahler conjecture holds in all dimensions;
• the plane cannot be unit-distance colored using five colors;
• counterexamples exist for Kaplansky’s zero-divisor conjecture;
• counterexamples exist for the Hadwiger conjecture.

Each of these would represent a substantial mathematical contribution if independently confirmed.

The key phrase is if independently confirmed.

At this scale, the verification problem becomes almost as interesting as the discovery problem itself.

🧪 The Model Behind the Results Remains Secret

#

One of the most conspicuous omissions is the identity of the model.

OpenAI refers to it only as an “internal frontier model.”

The company has not publicly disclosed its model name, architecture, or complete training methodology, saying that it is progressing toward a responsible release.

This is particularly interesting given another announcement from September concerning the Navier–Stokes problem.

OpenAI previously stated that an internal model significantly stronger than GPT-6 Astra had been trained beginning August 28. That system reportedly used approximately 10,000 parallel agents for 88 hours to work on the 3D incompressible Navier–Stokes equations and establish finite-time singularity formation under smooth external forcing.

GPT-6 Astra subsequently spent another 17 hours completing the Lean formalization.

The connection between that system and the model behind the 722-result catalog has not been fully disclosed.

What is clear is that OpenAI has been building systems designed not merely to answer mathematical questions, but to conduct extended mathematical research using large amounts of parallel computation.

⚠️ Why Mathematicians Are Concerned

#

The release arrives amid growing tension between frontier AI laboratories and the mathematical research community.

The dispute is not simply about whether AI can solve difficult problems.

It is about how mathematical discoveries should be released, verified, attributed, and integrated into the existing research ecosystem.

In August, OpenAI announced ten advances in mathematics and theoretical computer science attributed to its next-generation Astra model.

Soon afterward, mathematicians identified a problem with one of the reported results concerning a counterexample to the Connes rigidity conjecture: one of the groups used in the construction did not satisfy the required conditions.

That episode became an important warning.

A sophisticated-looking proof can fail because the system misunderstood a definition, overlooked a hypothesis, or proved a nearby statement rather than the intended one.

Anthropic subsequently disclosed progress from an unreleased model on problems related to the Riemann hypothesis, further intensifying the competition between AI laboratories.

According to reporting by WIRED, OpenAI privately convened roughly 40 mathematicians in August to discuss what should happen if AI systems begin surpassing human mathematicians on major research problems.

Some attendees reportedly said OpenAI indicated that its model had already solved hundreds of difficult open problems.

The mathematicians’ response was broadly consistent: do not merely announce the results on blogs and social media; publish them in a form that experts can actually inspect and build upon.

📚 The Publication Problem Is Becoming an AI Problem

#

The issue eventually became formalized into a broader discussion about standards for AI-generated mathematics.

On September 11, 25 Fields Medalists, including Terence Tao, signed an open statement warning about what they described as a serious misalignment between AI development incentives and mathematical research.

The concern is straightforward.

For mathematicians, solving famous problems is not merely about accumulating theorem counts. The process is also about developing concepts, understanding structures, establishing priority, connecting results to existing literature, and communicating ideas clearly enough that other researchers can extend them.

An AI system can potentially compress months or years of exploratory work into hours.

That creates an unprecedented publication bottleneck.

The mathematical community therefore began pushing for standards covering:

• disclosure of AI involvement;
• proper citation of prior work;
• peer review;
• formal verification where possible;
• preservation of research artifacts;
• reproducibility;
• attribution of human contributions;
• and sufficient explanation for researchers to understand the results.

The Leiden Statement, signed by more than 4,000 mathematicians, similarly called for stronger standards around AI-generated mathematical research.

🏛️ The IAS Recommendations

#

On September 29, the Independent Advisory Group on Mathematics and AI (AGMAI) at the Institute for Advanced Study published recommendations for AI-generated mathematical research.

The recommendations divide responsible publication into two stages.

Initial Release

#

AI laboratories should:

• search for and cite relevant existing work;
• write proofs using conventional mathematical standards;
• disclose the model and prompts;
• provide summaries of the model’s reasoning process;
• report time and compute costs;
• maximize formalization;
• preserve artifacts in a neutral repository;
• maintain persistent identifiers and revision histories;
• report failure rates on comparable problems.

Supporting Comprehension

#

The second stage focuses on what happens after publication.

Labs should support:

• conferences;
• workshops;
• expository material;
• broad access to AI systems;
• and mechanisms that allow the global mathematical community to understand and extend AI-generated results.

These recommendations are particularly relevant to the OpenAI release because the company explicitly cited them when announcing the catalog.

📊 How Well Does OpenAI’s Release Follow the Recommendations?

#

The answer is mixed.

OpenAI has made meaningful progress toward the proposed standards.

It provides:

• reasoning summaries;
• compute estimates;
• attempted-problem counts;
• Lean formalization for a substantial fraction of result families;
• version tracking;
• plans for workshops and conferences around major AI-generated discoveries.

But important gaps remain.

The most obvious is the missing model identity.

OpenAI has not disclosed the name of the system that generated the results.

The repository is also hosted under OpenAI’s own GitHub organization rather than a neutral third-party research archive. OpenAI has indicated that it is exploring alternative community-hosting arrangements.

The company has also acknowledged that future releases need improvement in areas including citation quality, mathematical exposition, and presentation.

So the release is neither a complete failure against community standards nor a perfect implementation.

It is better described as selective compliance with an emerging research norm.

🔍 Formal Verification Does Not Eliminate Human Review

#

The presence of hundreds of Lean formalizations is one of the strongest aspects of the release.

Formal proof assistants can make certain forms of verification dramatically more reliable.

But formalization does not eliminate the semantic problem.

Suppose an AI proves theorem \(A\) in Lean.

The proof assistant can establish that the encoded proof of \(A\) is logically valid.

It cannot, by itself, determine whether \(A\) is:

• the theorem the researchers intended;
• the strongest useful version of the theorem;
• correctly connected to the surrounding literature;
• based on appropriate definitions;
• or actually significant in the way the manuscript claims.

The earlier Connes-related episode illustrates why this matters.

A formal proof can be completely correct while the surrounding mathematical claim is wrong because the formalized statement does not capture an essential hypothesis.

The OpenAI repository’s Lean overview pages therefore distinguish between what has actually been formalized and what remains outside the formal verification boundary.

The quasi-Riemann family is a useful example: its core result may be formalized while later applications described in the manuscript remain outside that formalization.

This distinction is easy to miss but fundamental.

⏱️ The Real Bottleneck May No Longer Be Proof Generation

#

The most important implication of this release may therefore be broader than any individual theorem.

For decades, advanced mathematics has operated under a relatively simple constraint:

Expert mathematical reasoning is scarce.

Researchers spend months or years developing conjectures, constructing approaches, proving lemmas, discovering counterexamples, and finally writing papers.

If frontier AI systems can instead generate large numbers of plausible research results within hours, the bottleneck begins to move.

The scarce resource becomes expert attention.

Consider the scale of the OpenAI catalog.

There are 372 result families.

Even if each family required only one week of expert attention to understand and verify, a single mathematician would need more than seven years to process the entire collection.

And these are not textbook exercises.

Many of the claims involve areas where understanding the definitions, historical context, technical machinery, and surrounding literature can itself take months.

Formal verification can reduce the cost of checking logical correctness.

It does not automatically reduce the cost of understanding mathematical significance.

That distinction may define the next phase of AI-assisted mathematics.

🌍 From Automated Proof to Automated Discovery

#

The OpenAI release suggests a shift in what frontier mathematical AI is being optimized for.

Earlier systems were primarily evaluated through benchmarks:

• solve this equation;
• prove this theorem;
• answer this competition problem;
• generate this proof.

The new evaluation paradigm is different.

Give the system thousands of research questions.

Let it explore them autonomously.

Allocate substantial inference compute.

Allow it to build upon previous discoveries.

Then filter the resulting research into potentially meaningful result families.

This is much closer to an automated research pipeline than a conventional mathematical benchmark.

If the reported numbers are accurate, the system produced research candidates at a rate measured in hours rather than months or years.

That does not mean every output is correct.

It does mean the economics of mathematical exploration may be changing.

🚀 What the 722 Results Really Mean

#

It would be premature to conclude that AI has suddenly solved hundreds of major mathematical problems.

The appropriate interpretation is more nuanced.

The catalog demonstrates that a frontier AI system can apparently produce a large volume of highly sophisticated mathematical research candidates, some targeting problems of extraordinary difficulty.

A substantial portion has already been formalized.

Many claims remain unformalized.

And the most consequential question—whether the deepest results survive independent expert scrutiny—cannot be answered by the release itself.

The coming months will therefore be more important than the launch day.

If mathematicians verify even a small fraction of the strongest claims, the significance will be enormous.

If many collapse under expert review, that will also be valuable: it will reveal where current mathematical AI systems remain unreliable and where formalization needs to improve.

Either way, the experiment is informative.

🧭 The Bottleneck Is Shifting From “Proof” to “Comprehension”

#

The September Navier–Stokes announcement suggested that AI systems might be capable of attacking individual problems previously considered beyond practical machine reasoning.

This new catalog suggests something different.

The challenge is no longer only:

Can an AI prove a difficult theorem?

It is increasingly:

Can an AI generate enough potentially valuable mathematics that humans can no longer inspect all of it?

That changes the relationship between humans and mathematical AI.

When the supply of candidate theorems becomes abundant, the scarce resources become:

• mathematical judgment;
• verification;
• conceptual understanding;
• literature knowledge;
• exposition;
• attribution;
• and the ability to identify which results actually matter.

This is the deeper significance behind Sam Altman’s statement that we are entering a new era of discovery.

The 722 manuscripts are evidence of unprecedented mathematical output.

But output alone is not discovery.

Discovery happens when a result is verified, understood, connected to existing knowledge, and transformed into something that other humans can use.

The next frontier of AI mathematics may therefore not be generating more proofs.

It may be building systems—and research institutions—that help humanity comprehend thousands of proofs faster than any individual mathematician ever could.