OpenAI said an unreleased internal model produced over 100 long-standing mathematical results after training began August 28, 2026. The company described the training period as about a month. Its September 21 announcement came 24 days after the start, a shorter span than the stated training run.
The claim matters to mathematicians and the wider community because it raises questions about what counts as a result. Readers need to separate OpenAI’s reported progress from proof that others have checked and accepted. This account explains what the headline measures and how to assess the evidence.
The model that produced the manuscripts remains internal and unreleased. That distinction matters: the announcement describes a company claim, not an independent verdict. Treat 24 days as the gap between training start and announcement, not time spent on each problem.
Key Takeaways
- OpenAI reported more than 100 long-standing results across most areas of mathematics.
- The 24 days mark the time between training start and announcement.
- The model that produced the manuscripts has not been released.
- Manuscripts, result families, formal checks, and accepted solutions are different measures.
- Review each claim by its theorem and proof, not its headline count.
What Does OpenAI 100 Math Problems Solved 24 Days 2026 Mean?
The catalogue’s manuscript count and its count of result families measure different things. Neither count is a simple tally of unrelated problems that the internal model solved.
OpenAI posed approximately 4,000 problems in its evaluation, then grouped collected outputs using a significance threshold. Each evaluation result used an average of three hours of ChatGPT Pro thinking compute. The evaluation design and compute details describe those limits.
A manuscript count measures files, while a result-family count groups related mathematical claims.
What Is Listed on OpenAI’s Math GitHub?
OpenAI math problems on GitHub appear as manuscripts grouped into result families, not as a one-to-one list of solved problems. The OpenAI math repository includes manuscripts and supporting proof materials from the internal model.
A Lean formalization is a proof translated into Lean so software can check its logical steps. About 42% of top-line results have Lean formalizations. That leaves many results without a software-checked proof.
OpenAI cautions that unformalized results may have issues. The repository also describes its contents as raw research data, rather than polished papers. It released abridged reasoning summaries for 10 results, not every manuscript.
Does the Catalog Show 722 or 721 Manuscripts?
The current catalogue lists 721 manuscripts, not 722. A figure of 722 OpenAI math problems is not the catalogue’s current manuscript count. The distinction matters because a manuscript is a written output, not necessarily a separate problem or result.
Evaluation results and catalogue manuscripts also describe different stages. The evaluation posed roughly 4,000 problems, while the catalogue collects outputs that met OpenAI’s significance threshold. A single manuscript may discuss more than one related mathematical claim.
Are There 377 or 373 Result Families?
The current catalogue lists 373 result families, not 377. OpenAI’s math catalog groups related manuscripts into families, so the family total is smaller than the manuscript total.
Those figures help describe the scale and structure of the public mathematics collection. They do not settle how many distinct open problems have been solved or how strong each result is. A catalogue count describes what was collected, not the number of unrelated problems.
How Strong Is the Evidence Behind OpenAI’s Math Claims?
An AI-generated proof, a formally checked proof, and an accepted solution are not interchangeable claims. Independent review matters because a GitHub entry alone does not show that outside experts accept its result.
OpenAI reported results including the Hodge conjecture, another Millennium Prize Problem. It also proposed a solution to the Navier-Stokes existence and smoothness problem on September 8, 2026. OpenAI’s reported results are not the same as externally accepted solutions.
What Does the OpenAI Math GitHub Catalog Confirm?
The catalog confirms that OpenAI released mathematical manuscripts and proof materials. It does not, by itself, confirm that every theorem is correct or that every proof matches the problem as posed.
FrontierMath: Open Problems is a separate benchmark developed and solely owned by Epoch AI. It expanded to 50 problems on July 31, 2026. Its categories count only that benchmark’s problems, not OpenAI’s more-than-100 claim.
| FrontierMath category | AI-solved | Total |
|---|---|---|
| Moderately interesting | 6 | 22 |
| Solid result | 2 | 18 |
| Major advance | 1 | 6 |
| Breakthrough | 1 | 3 |
FrontierMath’s category counts separate AI-solved problems from the benchmark’s total. OpenAI is the only entity to have purchased access to its solution verifiers. Epoch AI funded and solely owns this open-problem benchmark, while OpenAI funded the original FrontierMath: Tiers 1-4.
A human-plus-AI label, added September 16, 2026, means AI helped in an important way but did not solve a problem on its own. The FrontierMath benchmark page uses that label separately from autonomous AI solutions.
What Happened With Navier-Stokes?
OpenAI said roughly 10,000 AI agents worked on the Navier-Stokes problem and reached a proposed solution after 88 hours. The effort involved nearly 3 million messages and 130 billion output tokens.
OpenAI said its work addressed two of the four statements required by the prize problem. It also said it did not plan to claim the Clay Mathematics Institute’s $1 million prize. The institute has not acknowledged any party for a correct solution.
Which Proofs Count as Verified?
A Lean formalization lets software check the logical steps that have been translated into Lean. That check does not alone establish independent acceptance or prove that a famous problem was solved exactly as posed.
For each claim, check these three points:
- Does the claimed theorem match the problem that was posed?
- Is a Lean formalization available?
- Have independent reviewers or the relevant prize institution accepted the result?
These checks separate mathematical progress from claims that still need review. For a prize-level claim, check the exact theorem and independent review status rather than relying on an agent count.
Who Can Shape OpenAI’s Math Research?
OpenAI’s advisory group gives mathematicians a channel to comment on how results are shared, not authority over the company’s research schedule. Its remit is limited: members can offer advice on releases, but cannot direct research.
OpenAI announced its Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study in September 2026. Nine mathematicians were named as initial members, including Fields Medalist Timothy Gowers. The unpaid group controls its own membership.
A count of solutions can miss the ideas that help people understand why a result matters.
That concern appeared in “A Severe Misalignment of AI in Mathematics,” an open letter signed by 25 Fields Medalists. The group can advise on sharing results, but has no decision-making power over OpenAI and cannot advise on research pace.
Readers can follow future AI and technology coverage through a newsletter with chosen topics. A useful measure of the group’s influence will be whether its public recommendations make future releases clearer.
Conclusion
The announcement signals rapid progress, but a headline count cannot settle the meaning or reliability of each result. Its importance depends on what the model proved and how others check the work.
A manuscript can offer a useful starting point for understanding a claim. Before relying on a particular theorem, read its manuscript and check whether its proof has been formalized.
FAQ
Why do 721 manuscripts correspond to 373 result families?
The 721 manuscripts correspond to 373 result families because related manuscripts are grouped together. A family can collect several written outputs about connected results, so each manuscript does not represent a separate family. The counts describe different units in the catalogue. Neither count alone tells readers how many unrelated problems OpenAI solved or how many claims experts have accepted.
Is the full reasoning behind every result public?
No, OpenAI released abridged reasoning summaries for 10 results, not a full summary for every catalogue entry. The repository includes manuscripts and supporting proof materials, but the reasoning summaries cover only that small set. A reader may need to examine each manuscript and its proof artifacts to understand how a particular claim was reached.
What does a Lean formalization verify, and what does it not establish?
A Lean formalization lets software check the logical steps translated into Lean. It does not by itself establish that independent mathematicians accept the result or that the theorem solves a famous problem exactly as posed. The formal check applies to the encoded proof and its stated assumptions, not to every broader claim made about it.
Has the Clay Mathematics Institute accepted OpenAI’s Navier-Stokes solution?
No, the Clay Mathematics Institute has not acknowledged any party for a correct solution. OpenAI proposed work on the Navier-Stokes existence and smoothness problem, but said its work addressed two of the four required statements. OpenAI also said it did not plan to claim the institute’s $1 million prize.
What does FrontierMath’s human-plus-AI label mean?
The human-plus-AI label means AI helped instrumentally but did not solve the problem autonomously. FrontierMath introduced the label on September 16, 2026, to distinguish these cases from problems solved independently by AI. The classification belongs to Epoch AI’s separate benchmark and should not be read as a category for OpenAI’s full catalogue.
