OpenAI published 372 mathematical results produced by an unreleased internal model on October 6, 2026, as a batch of manuscripts in a repository run by the company rather than as a paper in a journal. In OpenAI’s own description, they are new results on open problems in mathematics. Within four days the release had produced a second document: more than 100 public reactions collected by the math blog Proofs and Prompts, running from excitement about new proofs to grief about a profession that changed in a week.
What was actually released
The repository is the substance of the release, and OpenAI’s own catalog is specific about its size: 719 manuscripts grouped into 372 result families, each classified by mathematical discipline. A family can hold a principal result, companion arguments, consequences and alternative proofs, so the count of files is larger than the count of claimed results.
The part a reader can check without being a mathematician is the formal part. Lean is a programming language in which a proof can be checked by a computer, and OpenAI says it has formalized many of the proofs in Lean and will add more as it obtains them.
The repository also publishes ten abridged summaries of the model’s reasoning, estimates of the compute behind the work, and statistics on the problems it attempted. The procedure was mostly the same throughout: one unreleased model, about three hours of ChatGPT Pro thinking per result, and roughly 4,000 problems posed.
| What OpenAI published | Figure |
|---|---|
| Manuscripts in the catalog | 719 |
| Result families | 372 |
| Top-line results with a Lean formalization | 42% |
| Problems posed to the model | about 4,000 |
| Compute per result | about 3 hours of ChatGPT Pro thinking |
Two families were handled outside the fixed procedure: a zero-free region for the Riemann zeta function and the Hodge Conjecture for CM abelian varieties. OpenAI states that the writeup of the zero-free region for Re(s) greater than 11/12 was edited by a human for readability.
The numbers worth keeping
- The catalog is 719 manuscripts, or 372 result families.
- 42% of the top-line results carry a Lean formalization, the only part a reader can verify cheaply.
- OpenAI’s own repository warns that some of the unformalized results could still have issues.
What the claim rests on
The claim is that these are new results on open problems. What is verified today is the formalized subset, 42% of the top-line results. Everything else sits at different stages of verification, and OpenAI’s repository says plainly that some of the unformalized results could have issues. That is the most useful sentence in the release, because it tells a reader that 372 is a count of claims, not of established theorems.
Our arithmetic on OpenAI’s own numbers: about 4,000 problems were posed and 372 families were cataloged, which is roughly 9 results for every 100 problems the model was given. OpenAI does not publish a per-problem failure rate. The repository states only that grouping the output into families and requiring a level of significance produced the catalog, which leaves the selection rule inside the company.
The community response supplies the parts the release does not cover. In the answers collected by Proofs and Prompts on October 8, Henry Wilton wrote that OpenAI refuses to name the authors of its papers and does not submit them to journals, which in his reading presents mathematics as no longer a human pursuit.
If OpenAI wanted to destroy the mathematical community, this would be a great way to go about it.
Other responses turn the question of standards back on the community. Enrico Fatighenti wrote that he rejects a human paper he cannot read past the first page, yet the field is now expected to decipher and check papers from AI systems at its own cost. Giulio Tiozzo described the release as papers posted in a hurry, and reported taking one example, a group with 116 generators, and getting a smaller example with 12 generators in ten minutes.
What is not published
No named authors, no peer review, and no independent check. The results went to a repository the company controls, with issues and pull requests disabled, so the community has nowhere inside the release to post corrections. The model that produced the results is unreleased, so the run cannot be reproduced from outside.
On this page
What it changes for anyone paying for frontier AI
The release is not a product, and the model behind it cannot be bought yet. Its value to a buyer is a signal about what a frontier model can be trusted with in a field where a wrong answer is not a matter of taste.
The first signal is the yield. A laboratory that poses 4,000 problems and publishes 372 result families is publishing its evaluation byproducts, and the ratio is what matters for planning: about nine cataloged results per hundred problems tried, and only 42% of those machine-checked. If you build on AI-assisted technical work, that ratio is the shape of the pipeline as the supplier itself describes it.
The second is the cost of verification. Fatighenti and Johannes Schmitt make the same point: the papers exist, the readers do not, and the burden of checking falls on whoever cares about the result. For a team already using AI for research or engineering, the practical rule is to budget for the check, not only the generation: formal verification where the domain allows it, and human review time in the plan where it does not.
The third is provenance. The papers carry no author names, so citing one means citing a repository rather than a person. That is a weaker footing than a journal, and worth knowing before a result ships.
The Neural Minute take
The number to remember from this release is not 372. It is 42%.
That is the fraction OpenAI itself reports as machine-checked, and the same repository states that some results outside it could have issues. A supplier that publishes 719 manuscripts and then says in its own repository that part of the pile may be wrong is more honest there than in its headline. The headline is a claim; the formalized 42% is a fact. Treat the machine-checked subset as the deliverable and the rest as a research program, and ask any vendor making a comparable claim which fraction has been checked, and by what.
Two further points matter for anyone building on this. The model is unreleased, so the work is reproducible only inside OpenAI, which makes the release closer to a paper than to a tool you can run. And the reaction arrived within two days, faster than the papers can be read, which is itself evidence that the field has no process for output at this volume and knows it.
Our reservation is about the count. The 372 families, the 719 manuscripts and the 42% are all figures OpenAI reports about its own work, with no external check and no named author behind any paper. They may well be right. They are not independently verified yet, and the release, to its credit, says so.
Sources
- Sharing AI progress in mathematics, openai.com, published October 6, 2026: https://openai.com/index/sharing-ai-progress-in-mathematics/
- The catalog figures used here, 719 manuscripts, 372 result families, 42% of top-line results formalized, about 4,000 problems posed and about three hours of ChatGPT Pro thinking per result, are published in the repository OpenAI runs under its own GitHub organization, read on October 10, 2026
- The reactions quoted here come from the collection published by the math blog Proofs and Prompts on October 8, 2026