On October 6, 2026, OpenAI did something no AI lab had done before at this scale: it opened a public GitHub repository — github.com/openai/math — and uploaded 722 mathematics manuscripts in a single drop, all produced by an internal research model the company has not named and has not released to anyone outside the company. The manuscripts are grouped into 372 "families" of related results, spanning roughly 17 subject areas from number theory and algebraic geometry to theoretical computer science, mathematical physics, and formal logic (Sources: Engadget, mlq.ai).
The scale alone is unusual — most individual mathematicians publish a handful of papers a year, and a prolific research group might manage a few dozen. OpenAI published more manuscripts in one day than most university math departments produce in a decade. That volume is exactly why the release has become as much a story about how mathematics gets verified as it is about what the model actually found.
Quick summary: On October 6, 2026, OpenAI published 722 math manuscripts — grouped into 372 result families covering roughly 4,000 problems it says the model attempted, across 17 subject areas — to a public GitHub repository, produced by an unnamed, unreleased internal model. Each manuscript reportedly used about three hours of ChatGPT Pro-level compute. The repository includes supporting proof artifacts, Lean formalizations of the main results for 162 of the 722 manuscripts, and 10 abridged summaries of the model's reasoning. Headline results include a tightened partial bound related to the Riemann hypothesis (a 'zero-free' region pushed from Re(s) > 1.0 to Re(s) > 0.875 — still far from the conjecture's actual Re(s) = 0.5 target), a reported complete formula for a restricted case of the Birch and Swinnerton-Dyer conjecture, and results on restricted cases of the Hodge conjecture — two of which were withdrawn the next day over a sign error (see our follow-up post on the backlash). None of this solves a Millennium Prize Problem outright; each result applies to a narrower, specific case of a much harder general conjecture.
What's actually sitting in the GitHub repo
The repository itself is more than a pile of PDFs. Alongside the 722 manuscripts, OpenAI included supporting proof artifacts and, for the main results of 162 manuscripts, formal verifications written in Lean — a proof assistant that checks a mathematical argument step by step against formal logic rather than relying on a human reviewer reading it for mistakes. OpenAI also published 10 abridged summaries explaining the model's reasoning process on a handful of the higher-profile results, an apparent attempt to make at least some of the work legible to human reviewers who can't read all 722 full manuscripts (Source: mindstudio.ai).
Each individual manuscript reportedly took the model roughly the equivalent of three hours of ChatGPT Pro-level thinking compute to produce — not three hours of wall-clock time necessarily, but three hours' worth of the heaviest reasoning tier OpenAI sells to subscribers. Multiply that by 722 and the compute bill behind this single release is substantial, even before counting the much larger number of attempts that presumably didn't produce a usable result (Source: mindstudio.ai).
The headline results — and what they don't claim
Start with the Riemann hypothesis, probably the single most famous unsolved problem in mathematics (its Clay Institute bounty is $1 million, as one of seven original Millennium Prize Problems — one of which, the Poincaré conjecture, has since been solved). The real Riemann hypothesis says every "non-trivial zero" of the Riemann zeta function has real part exactly 0.5. Nobody has proven that. What OpenAI's model reportedly did is narrow a related "zero-free region" — a boundary past which zeros provably can't exist — from the previous best-known bound of Re(s) > 1.0 down to Re(s) > 0.875. That's real, incremental progress on a specific sub-question some coverage of the release nicknamed the "quasi-Riemann hypothesis" — not a proof of the Riemann hypothesis itself, and it doesn't claim to be (Sources: BigGo Finance, mindstudio.ai).
On the Birch and Swinnerton-Dyer conjecture — another of the seven Millennium Prize Problems, concerning elliptic curves — the model reportedly established a complete formula for a restricted class of cases where a mathematical object called the Selmer group has "corank" zero or one. If that result holds up to review, it would be a meaningful advance in a specific corner of a field where the last major breakthroughs date back to the 1980s. It is, again, a result about a narrower special case, not the general conjecture (Source: mindstudio.ai).
The model's Hodge conjecture results — covering CM abelian varieties and products of K3 surfaces, rather than the Hodge conjecture in its full generality — turned out to be the release's weakest point. Less than 24 hours after the October 6 upload, OpenAI disclosed that two of those manuscripts, plus a third related one, contained a sign error in a supporting construction and had to be withdrawn, alongside 14 lower-stakes revisions elsewhere in the batch. We cover that retraction, and the backlash it triggered among mathematicians, in a separate follow-up post.
The one-sentence version worth repeating: none of these results solves a Millennium Prize Problem. Each one chips away at a narrower, specially-constructed case of a much harder general conjecture — genuinely useful mathematics if it survives review, but not the headline-grabbing 'AI solves century-old math problem' some early social-media reaction implied.
Not OpenAI's first unsupervised math moment
This wasn't the company's first attempt at autonomous mathematical research — in May 2026, OpenAI said an internal model had disproved a 1946 conjecture by Paul Erdős concerning unit distances, a smaller, single-result claim that drew far less attention than this 722-manuscript release has. The October drop is a difference of scale and ambition: instead of one result, OpenAI pointed the model at what it describes as roughly 4,000 open problems across 17 fields and published whatever came out the other end, verified or not (Source: Startup Fortune).
OpenAI itself has said the model resolved something like 90 of roughly 500 problems from a list of especially hard, tracked open questions it was working against — a claim mathematicians are actively in the process of checking rather than one that's been independently confirmed. Given how fresh the release is and how specialized each of the 17 subject areas is, that verification is expected to take a long time: there simply aren't many people in the world qualified to check some of these proofs line by line (Source: Startup Fortune).
Who this matters for
- Researchers in number theory, algebraic geometry, and theoretical computer science who want to check whether a specific one of the 372 result families actually advances a problem they work on.
- Anyone following the 'AI for science' narrative broadly — this is one of the largest single tests yet of whether a frontier model can produce genuinely new, correct mathematics rather than restating known results.
- Students and educators trying to understand what 'AI does math now' actually means in practice, versus what headlines imply.
- Not, yet, anyone looking for a settled verdict — real verification of specialist mathematics by a small number of qualified reviewers takes months to years, not days.
01Did OpenAI's model solve the Riemann hypothesis?
No. It reportedly narrowed a related 'zero-free region' bound from Re(s) > 1.0 to Re(s) > 0.875 — real incremental progress on a specific sub-question, but nowhere near proving the actual conjecture, which claims all non-trivial zeros sit exactly at Re(s) = 0.5.
02Did any of OpenAI's 722 manuscripts get withdrawn?
Yes. On October 7, one day after the release, OpenAI withdrew three manuscripts related to the Hodge conjecture after a sign error was found in a supporting construction, and issued 14 further revisions across the rest of the batch.
03What model produced these manuscripts?
OpenAI has not named or released the internal model responsible. It is described only as an unreleased research system, distinct from any publicly available ChatGPT model.
04How much of this has been independently verified?
Very little, as of this release. OpenAI included Lean-formalized main results for 162 of the 722 manuscripts and 10 abridged reasoning summaries to help reviewers, but with 722 manuscripts across 17 specialized subject areas, full independent verification by the small pool of qualified mathematicians in each field is expected to take a long time.
05Is this the same as OpenAI's earlier Erdős problem claim?
No — in May 2026, OpenAI said an internal model disproved a single 1946 Erdős conjecture about unit distances. The October release is a much larger, 722-manuscript batch attacking roughly 4,000 problems across 17 fields at once.
Whatever your take on AI-generated mathematics, it's part of the same wave of model capability that's reshaping everyday tools too — coding assistants, research helpers, and chatbots all draw on the same underlying progress. LumiChats gives you access to 40+ premium AI models, including the latest from OpenAI, Anthropic, Google, and xAI, side by side in one place, so you can see how today's frontier models actually reason through a hard problem before deciding which one to trust with yours.
