On Tuesday evening, 6 October 2026, OpenAI published Sharing AI progress in mathematics and opened a public GitHub repo, openai/math. The drop is huge: OpenAI says the catalogue holds 722 manuscripts organized into 372 families — each family a principal result plus companions, consequences, or alternative proofs, classed by discipline.
OpenAI claims those results resolve or make substantial progress on major open questions in mathematics and theoretical computer science. Highlighted claims include a solution to the four-dimensional Kakeya conjecture, improvements to important algorithms, and progress toward the Riemann hypothesis — not a full proof of RH, but a zero-free region for the zeta function. The Verge and Scientific American both covered the release within hours.
What you cannot do today: run the model yourself. OpenAI has not released it. Readers should treat every result as an OpenAI claim until mathematicians — and Lean, where available — finish checking the work.
What’s in the Drop — 722 Manuscripts, 372 Families
Per OpenAI’s README, the vast majority of results came from one fixed procedure on an unreleased internal model. On average, each result used about three hours of ChatGPT Pro thinking compute. The model was posed roughly 4,000 problems; aggregating significant outputs produced the 372 families.
Exceptions to that fixed procedure include work on a zero-free region for the Riemann zeta function and a proof of the Hodge Conjecture for CM abelian varieties. OpenAI says the write-up for the Re(s) > 11/12 zero-free region was human-edited for readability.
Ten published reasoning summaries cover topics such as the irrationality exponent of π, the Mahler conjectures, Kaplansky’s direct-finiteness conjecture in characteristic two, and free group factor isomorphism. Peer review and community parsing will take months — Scientific American notes it will take mathematicians a long time to decide which proofs contain novel ideas and which are mostly mash-ups of existing techniques.
Verification — Lean Helps, but Not Everything Is Checked
OpenAI is explicit: not all results have accompanying Lean formalizations. Lean is a proof-checking language; Scientific American notes that Lean-checked results are all but certain to be correct. OpenAI also warns that some of the unformalized results could have issues, and says it will fix problems, add formalizations, and keep version history.
That split matters. A Lean-verified claim is in a different trust bucket from an unformalized manuscript dumped into a repo overnight. OpenAI itself is asking the field to treat the collection as a living catalogue — not a finished, peer-reviewed journal issue.
The Model You Can’t Use — Same Lineage as the Navier–Stokes Claim
The model is unnamed and not public. OpenAI says it is the same internal model behind its 8 September Navier–Stokes claim — a claim OpenAI said was produced by a large agentic swarm and for which OpenAI said it will not claim the Millennium Prize. That earlier claim remains contested and unsettled — not settled “solved” fact.
An OpenAI spokesperson told Scientific American that almost every result in the new drop came from a single prompt to a single agent — a striking contrast with the Navier–Stokes swarm — though the spokesperson also said some results may have taken multiple attempts. Until the model ships, nobody outside OpenAI can replicate the procedure.
Why Mathematicians Are Uneasy
After the Navier–Stokes controversy, OpenAI assembled an independent advisory group (AGMAI) whose recommendations — disclosed model, exact prompt, and compute per result; deposit in repos no lab controls — OpenAI is not fully following. OpenAI disclosed average compute and some stats, not prompts, and told Scientific American it is not bound by the recommendations while saying it takes them seriously and is working to release the model responsibly.
Andrew Sutherland of MIT told Scientific American: “Until and unless they release the model and people can replicate their results, I think you should treat any claims about one-shotting problems with a single agent as unverified… We should ask for receipts.” Daniel Litt of the University of Toronto took a more open stance: “If we want to know the answers to these math questions, I see no reason why we should ask the company to keep them secret from us… To me, it’s going to be a good thing for mathematics.”
On Hacker News, number theorist JoshuaZ called one quasi-Riemann-related result “a massive big deal” that would likely be Fields Medal territory for a human — while pushing back on overheated “biggest in 200 years” framing. Scientific American paraphrased Terence Tao as criticizing the “insane” pace of frontier-lab AI math dumps. OpenAI’s spokesperson also told Scientific American that many of the newly released results are not yet understood by its own mathematicians.
JamoraquAI Take
The number isn’t the story — the receipts are. Seven hundred twenty-two papers from a model nobody outside OpenAI can touch is a flex until mathematicians check the work. Trust the Lean-verified slice first; treat unformalized claims as open tickets, not settled theorems. Progress on a zero-free region is not “AI solved the Riemann hypothesis.”
What should matter to everyone else: three hours of “Pro thinking” per result means research-grade reasoning is getting cheap fast. When that model ships, the bottleneck in a lot of fields stops being “who can solve it” and becomes “who can check it.” Learn to verify, not just prompt.
Sources
- OpenAI — Sharing AI progress in mathematics (6 October 2026)
- GitHub — openai/math (README: 722 manuscripts / 372 families)
- The Verge — OpenAI math release on GitHub (Robert Hart, 6 October 2026)
- Scientific American — OpenAI unleashes hundreds more math results (Joseph Howlett, 6 October 2026)
- Unite.AI — OpenAI releases 722 math manuscripts from an unreleased AI model
- Hacker News — discussion thread (JoshuaZ comment)



