OpenAI’s Unreleased Model Claims Hundreds of Mathematical Results, but Researchers Want Evidence

Table of Contents

• You might want to know


• Main Topic


• Key Insights Table


• Afterwards...

You might want to know

• What does OpenAI’s release of 722 mathematical manuscripts demonstrate—and what does it not yet prove?


• Why are mathematicians asking for the model, prompts, and independently checkable evidence?

Main Topic

OpenAI has published 722 mathematical manuscripts on GitHub, describing them as outputs from an internal AI model that the company has not yet released. An OpenAI spokesperson said almost all of the work came from a single prompt given to a single AI agent, although some results may have required multiple attempts. If independently confirmed, the work could represent a significant development in the use of AI for mathematical research. For now, however, the model and the exact process behind the results are not publicly available, leaving researchers unable to fully reproduce the claims.

The collection is organized into 372 families of related results. A family may include a central theorem alongside supporting arguments, consequences, or alternative proofs. As a result, 722 is the number of manuscripts—not a count of 722 separate mathematical problems solved. OpenAI says it gave the model roughly 4,000 problems and selected outputs it considered significant enough to publish. That distinction matters: a large manuscript collection may contain many related pieces of work, and the number of documents alone does not establish how many independent problems have been resolved.

OpenAI reports that an average result used computing equivalent to roughly three hours of ChatGPT Pro thinking compute. The source article contrasts this with a Navier-Stokes claim from last month, which involved 10,000 coordinating agents working for 88 hours. These descriptions suggest that the two efforts used quite different approaches and amounts of computing. They do not, by themselves, establish the correctness or importance of either set of mathematical claims.

MIT mathematician Andrew Sutherland urged caution. He told Scientific American that claims about solving problems in one prompt with a single agent should be treated as unverified until the model is released and other researchers can replicate the results. His request for “receipts” reflects a basic principle of scholarly evaluation: researchers need enough information to examine the methods, check the arguments, and test whether the reported conclusions follow.

OpenAI released abridged reasoning summaries for 10 results. A repository catalog indicates that just 162 of the 722 papers have a main result formalized in Lean, a programming language and proof assistant that can mechanically check logical steps. This represents about 22% of the collection. OpenAI has acknowledged that not all manuscripts have Lean formalizations and warned that “some of the unformalized results could have issues.” The company has said it plans to add formalizations as it obtains them.

A successful Lean check is useful evidence, but it is not a complete verdict on a mathematical claim. It confirms that a proof follows from the statement encoded in Lean. It does not establish that the formalized statement accurately captures the original problem, that the result is new, or that it is important. Mathematicians still need to assess those questions, as well as whether the argument is meaningful and whether its assumptions match the intended context.

That distinction has drawn attention to the character of some of the arguments. In a post dated October 7, 2026, Dmitry Rybin described an attempt to read an OpenAI proof concerning whether the chromatic number of the plane is greater than or equal to 6. He said the argument was difficult to understand and questioned a step connecting a coloring to a “weakly measurable” coloring. The post is an individual researcher’s reaction, not a general assessment of the entire collection, but it illustrates a wider concern: an argument may be difficult for people to interpret even when it is presented as a proof.

Other researchers have raised practical questions about how the community can scrutinize and improve the work. In a post also dated October 7, 2026, Keith Adler criticized the repository for having its Issues feature turned off and for not accepting pull requests. He argued that a project presenting 722 manuscripts and inviting Lean formalizations should provide a clear place for researchers to submit comments and contributions. These observations point to the importance of collaboration infrastructure: reproducibility depends not only on publishing results, but also on enabling people to report errors, propose corrections, and share independent checks.

The Institute for Advanced Study in Princeton, New Jersey, emphasized the human dimension in a statement. It said that AI can now produce mathematical arguments that the person prompting it may not be able to understand, verify, or take responsibility for. The institute argued that human understanding remains paramount and asked how a responsible approach to mathematical publication can preserve it. This concern is not simply about whether an AI system can produce a valid proof. It is also about who can explain the reasoning, evaluate its significance, and take responsibility for the claims being published.

Mathematicians have also offered a more positive assessment of the collection. Professor Abhishek Saha called the release “a very big day for mathematics.” He nevertheless distinguished between exceptional advances within an existing research program and surprising breakthroughs. In his view, most of the results appear interesting without necessarily being impossible or transformative. The source article identifies one result—the Quasi-Riemann Hypothesis—as the collection’s possible standout. That characterization remains a matter for mathematical experts to assess, rather than a conclusion established simply by the size of the release.

Transparency is another point of debate. An advisory group at the Institute for Advanced Study recommended on September 29 that releases of this kind include the model name, prompts, a summarized chain of thought, the time taken, and the compute cost for every result. OpenAI has shared average compute information and 10 reasoning summaries, but not the prompts. The company says it is still working on a responsible way to release the model. Without those details, independent researchers have fewer means to reproduce the work or determine how much a prompt, repeated attempts, or other aspects of the process contributed to each result.

Not all researchers agree that withholding the model or answers is the right approach. Daniel Litt, a mathematician at the University of Toronto, argued that there is no reason to ask the company to keep the answers to these mathematical questions secret. The disagreement reflects a broader tension: early access may help researchers inspect claims, while a controlled release may be intended to address concerns about responsible deployment. Whatever approach is chosen, clarity about what information is available—and what remains unavailable—will be important for assessing the evidence fairly.

A comparison with another recent effort helps clarify what is distinctive about OpenAI’s announcement. Last month, Anthropic posted all 13 million lines of its Lean-checked proof of Fermat’s Last Theorem on GitHub. That work formalized a theorem Andrew Wiles published in 1995; it did not claim new mathematical results. OpenAI’s release, by contrast, presents a large set of results as outputs from an unreleased model, with only 162 manuscripts currently accompanied by a Lean-formalized main result. The two cases therefore differ both in the kind of claim being made and in the amount of material available for direct inspection.

The most careful interpretation is neither to dismiss the collection nor to treat its claims as settled. The manuscripts could contain valuable mathematics, but their number does not establish the number of independent problems solved, and a proof assistant check does not answer every question about correctness, novelty, or significance. Researchers will need to inspect the arguments, test the formalizations, compare results with existing literature, and assess the model’s performance under reproducible conditions.

Key Insights Table

Afterwards...

The next stage of AI-assisted mathematics will depend on more than producing a large volume of plausible-looking proofs. Researchers will need methods that make results independently reproducible and that clearly distinguish between a model’s generated argument, a mechanically checked proof, and a result judged mathematically significant by experts. Publishing prompts and relevant process details, where responsible, could help clarify how individual outputs were obtained. Open repositories with accessible ways to report issues and submit formalizations could also make scrutiny more effective.

Further work is needed on proof assistants and tools that help people move between informal mathematical writing and formal statements. Better systems could flag gaps, surface assumptions, and make complex arguments easier to inspect without treating automated verification as a substitute for expert judgment. Evaluation practices should also examine whether a claimed result is genuinely new and whether its formal version faithfully represents the original question.

Ultimately, the most promising direction is a collaborative one: AI systems may help explore ideas and construct candidate arguments, while mathematicians provide interpretation, verification, and accountability. Progress will be measured not only by what models can produce, but by how clearly people can understand, test, and responsibly build on it.