OpenAI Dropped 722 AI-Written Math Papers, and the Community Is Demanding Receipts

An AI model wrote 722 mathematical manuscripts in one batch, among them the sharpest attack on the Riemann hypothesis in 160 years, and the people who should be throwing a party are furious. On October 6, OpenAI opened the public repository openai/math on GitHub: 722 manuscripts organized into 372 result families, plus Lean formalizations covering 235 of those families, which means a computer can check the proof steps line by line instead of asking you to trust an AI. Sam Altman posted on X that we are entering a new era of discovery. The mathematical community answered with mafia comparisons, an open letter, formal recommendations from an elite advisory group that begin with the word stop, and an essay from an xAI co-founder explaining why mathematics will never be the same either way. I am walking through the whole chain because it dramatizes something I deal with every day while running an agent setup: the difference between what an agent claims and what a checked result is. The trigger for this post is a video by AI Revolution that assembles the storyline; every core number below was checked against the official OpenAI announcement and the repository itself.

First the facts. The manuscripts come from an internal frontier model at OpenAI that nobody outside the company can run. According to OpenAI, the team pointed the model at roughly 4,000 real open problems and kept only the significant results. The headline everyone looked at first was Riemann. The Riemann hypothesis states, since 1859, that the nontrivial zeros of the zeta function all sit on one straight line; hundreds of theorems about prime numbers lean on that assumption, yet nobody managed to prove it in 160 years. The model has now proven, for an entire family of related functions, a zero-free region where the feared outliers cannot hide, and it also ruled out a so-called Siegel zero for that family, the phantom that costs analytic number theorists decades of sleep. That is not the hypothesis itself, but if it holds, it is the biggest step in this area in half a century, and the steps are machine-checked in Lean. The same pattern repeats across the repository: the unique games theorem, proving Subhash Khot's 2002 conjecture outright along with its 2008 consequences; partial results on the Hodge conjecture; the Kakeya problem in three and four dimensions; a version of Hilbert's tenth problem; and the irrationality of Catalan's constant. What did not happen: a complete proof of the Riemann hypothesis or of the Navier-Stokes equations.

The model is the successor of the one that hit the Navier-Stokes millennium problem last month, and OpenAI says it is far stronger than GPT-6 Astra. The effort figure is what makes professionals rub their eyes. For the millennium result, OpenAI reported roughly 10,000 agents running for 88 hours; this time, almost every result reportedly came from a single agent with a single prompt, averaging about three hours of ChatGPT Pro thinking. Special treatment was limited to the Riemann and Hodge write-ups, with humans polishing the Riemann paper. OpenAI published ten reasoning summaries, the average compute, and the number of attempted problems, but the prompts themselves are missing, and so are per-result costs. Andrew Sutherland of MIT accordingly said the one-agent claims stay unverified until the model ships and the results can be reproduced; his public framing was that we should demand to see the receipts. One more detail worth watching: the repository is alive. Manuscripts and Lean formalizations were pushed forward several times within the first days, which tells you that errors in the unchecked portions are real and are being fixed version by version, exactly the lifecycle that Lean's revision protocol makes visible.

Why is nobody celebrating? Because the relationship between the lab and the discipline is broken, and the breakage has history. In August, OpenAI reportedly gathered around 40 leading mathematicians behind closed doors and asked what the discipline would do if AI completely overtook it; attendees say the company hinted it had already cracked more than 100 open problems. That was officially confirmed on September 21, along with a promise not to dump everything at once. The mathematicians, according to their own accounts, begged OpenAI to behave like a publisher rather than an influencer and to release papers humans had time to check. Bryna Kra of Northwestern University later summarized that the room felt extreme excitement and extreme fear at the same time, and that these opinions were clearly ignored. On October 6, all 722 manuscripts landed at once. An OpenAI spokesperson said the company is not aware of any such promise. Then there are the open wounds: an open letter signed by 25 mathematicians in September accuses the labs of using humanity's hardest problems as a scoreboard for their models. The same nerve is hit by the Navier-Stokes fight: Tristan Buckmaster of NYU says he was working on a key piece of that problem with Levent Alpoge of Anthropic, using OpenAI's own tools, and that OpenAI, spurred by a rumor that another team was close, threw thousands of agents at the problem to get there first. Lines that OpenAI researcher Sebastian Bubeck allegedly said during name negotiations are denied by Bubeck himself; I treat those details with caution because they rest on Wired reporting I have not read in full. Néstor Guillen, a visiting professor at NYU, puts the mood in one sentence: the fear among his colleagues is about these companies and how much power is piling up in one place, not about AI itself, which is exactly where the mafia comparisons come from. And with OpenAI and Anthropic both preparing to go public, many read the mass release as showing off for investors.

The body OpenAI itself points to is the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study in Princeton, and its members include Fields medalists Timothy Gowers, Martin Hairer, and Edward Witten. After hearing from more than 600 mathematicians, the group published recommendations on September 29 that open with a radical line: labs should stop testing big math problems on private models entirely. The rest reads like an etiquette code for agent operators. Every result should disclose which model was used, which prompts, how the reasoning went, how long it took, and what it cost. Proofs should be machine-checked, in Lean or something like it. Results should be stored somewhere no AI company controls. Marketing stays out. Large batches should document how the problems were selected and how many attempts failed. Funding for conferences, summer schools, and workshops should flow through nonprofits that decide where the money goes. And mathematicians should get access to publicly available models. OpenAI has not adopted the guideline set as binding; the company says it is not bound by them, argues that testing internal models is how it builds science tools faster, promises clearer papers and workshop funding, and continues looking for a community-run home for the papers, which for now sit in an OpenAI-operated repository. A spokesperson even admitted that OpenAI's own mathematicians have not fully understood many of these results yet.

I have not verified any of the 722 claims myself, and I say that not as a disclaimer but as the method. First lesson: what makes a result valuable is formalizable checking. An agent can assert a million words, but the value of an assertion depends on whether a mechanism can recount the steps without listening to anyone; Lean plays exactly the role for mathematics that the readback plays in my agent setup. I do not trust an ok: true from an API, I read the result back. The community is building precisely this backbone on its own initiative, with registries for machine-checked proofs and a Leiden Declaration signed by more than 4,000 mathematicians. Second lesson: sovereignty shows up in where things are stored, not in which model is smartest. The real dispute in this whole affair is not whether the results are correct but who controls the storage, who discloses costs, and whose rules apply. Translated into my operation, those are the same questions I ask about any tool that sells me compute as a subscription: I keep data local, I measure in my own dashboard, and I trust numbers from the user's dashboard, not from marketing.

The sharpest pro position comes from Christian Szegedy, one of the twelve founders of xAI, who previously led work on Inception, on adversarial examples, and co-created batch normalization, and who predicted in 2019 that AI would become a superhuman mathematician by 2029 by learning to turn human mathematics into machine-checkable proofs. In 2024 he moved that prediction up by three years and bet François Chollet, the creator of Keras, and others that it would happen before June 2026. The surrounding context this year: OpenAI disproved an 80-year-old geometry conjecture in May; Anthropic researchers used Claude Fable 5 in July to knock down a conjecture from 1939; and on September 4 an internal Anthropic model produced a computer-checked proof of Fermat's Last Theorem, which took humans over 350 years, in just 11 days. His essay on X asks why mathematics was a jungle expedition and what happens when airplanes arrive. The Wright brothers' first flight lasted 12 seconds, shorter than the wingspan of an airliner, and that, says Szegedy, is where AI stands in mathematics right now. The climbers will insist that the rocks the planes land on are uninteresting and that the real peaks will never be reached; he calls that static thinking. Once flying gets cheap, anyone can charter a plane, satellites will map every treasure, and the old explorers will blame the big companies for ruining the jungle, which he says is fair but will change nothing. His closing line is four words long: don't bet against airplanes. I read the essay not as a forecast but as a description of what is actually happening in the repository dispute: the discipline is not negotiating about mathematics, it is negotiating about the infrastructure of truth.

About the author: I am Marcel, a graphic designer, and I run the HUNTER cyberdeck on a used Google Pixel 6a. I orchestrate my agent family with Hermes, pi, and herdr, and I write these posts from live operation. Last fact-checked on October 8, 2026.

Sources: OpenAI, Sharing AI progress in mathematics, official announcement of the release with reasoning summaries, compute figures, and cooperation with the advisory group, retrieved October 8, 2026. GitHub, openai/math, repository with 722 manuscripts and Lean formalizations, star and license figures from the repo index, retrieved October 8, 2026. Advisory Group on Mathematics and Artificial Intelligence, general recommendations of September 29, 2026, as linked in the OpenAI text. Scientific American, OpenAI unleashes hundreds more math results upon a field already in shock, statements by Terence Tao and Daniel Litt, retrieved October 8, 2026. latent.space, AINews on quasi-Riemann with community reactions including the Alpoge quote, retrieved October 8, 2026. YouTube, AI Revolution, OpenAI Just Broke Math With Its Most Powerful AI Yet, trigger video, reviewed October 8, 2026.