OpenAI’s 722 AI Math Manuscripts Claim Quasi-Riemann Breakthrough

An Overnight Dump of 722 Manuscripts

In a move that flouted the usual rhythms of mathematical publishing, OpenAI dropped 722 research manuscripts into a public GitHub repository on October 6, 2026, covering 372 distinct result families across number theory, geometry, theoretical computer science, algebra, mathematical physics and differential equations13. The papers came from an internal frontier model the company has not released, evaluated on roughly 4,000 open research problems, with an average of about three hours of ChatGPT Pro-equivalent thinking compute per result24. Sam Altman framed the release as "a new era of discovery"2. Nothing in the collection has been peer reviewed, and OpenAI itself cautions that unformalized results may contain errors4.

The headline item is Result 003, which OpenAI brands the "quasi-Riemann hypothesis": a claimed proof that the Riemann zeta function and all Dirichlet L-functions have no zeros in the half-plane where the real part of s exceeds 7/8, with a companion manuscript giving an alternate argument for the weaker 11/12 boundary37. A third manuscript in the family claims a uniform exclusion of Landau–Siegel zeros, the long-standing ghost of analytic number theory6. If it survives scrutiny, commentators have placed the result somewhere between a Fields Medal-level achievement and the biggest number-theory result in two centuries2.

What the Quasi-Riemann Claim Actually Says

The classical Riemann hypothesis demands that every nontrivial zeta zero sit exactly on the line with real part 1/2. The quasi-Riemann statement is weaker but still extraordinary: it asks for some fixed constant below 1 such that no zeros exist to its right, at any height5. Known zero-free regions historically shrink as you move up the critical strip; a fixed boundary at 7/8 that holds uniformly would be an unprecedented advance in analytic number theory within the past half-century1.

The reported proof machinery is classical in flavor — Poisson summation, an iterated "dual mean square" estimate, and extensive theta-function calculations — described by analysts as deep technical work rather than brute-force search7. At its core, the argument measures the same character-sum two ways: a direct bound on one hand, and a Mellin integral involving the reciprocal of an L-function on the other; the claimed power savings let the reciprocal be analytically continued past a hypothetical rightmost zero, producing a contradiction3. The manuscript runs 199 pages10.

Beyond Riemann-adjacent work, the catalogue claims the Unique Games Conjecture, the free group factor isomorphism problem, both Mahler conjectures, the Hodge conjecture for CM abelian varieties, an exact irrationality exponent for pi, 4D Kakeya results, large-data global regularity for the relativistic Vlasov–Maxwell system, and a uniqueness result for the elastic inverse problem open in 3D since 1994352. Notably, an estimated 20% of the results are disproofs or counterexamples, undercutting the claim that AI math is merely pattern-matching toward positive results2.

Verification, Lean, and the Limits of Formal Proof

OpenAI's release is not just PDFs. Some 235 of the 372 result families link to Lean formalizations, with a formalization catalogue listing 162 papers whose main result is machine-checked, and challenges runnable through the Lean Comparator tool43. This is the scientific-computing angle that matters most: machine-verifiable proof artifacts change the economics of trust in mathematics. A Lean certificate can be independently reproduced by anyone with the specified toolchain versions, sidestepping some — though not all — of the traditional peer-review bottleneck10.

But the caveats are real and OpenAI states them plainly. "Many, but not all" manuscripts have been formalized; unformalized ones "could have issues," and corrections will be recorded as new versions6. The zero-free-region write-up fell outside the model's fixed generation procedure and was edited by humans for readability47. Coverage gaps between the formal theorem and the full manuscript are easy to miss but fundamental — the Lean proof may verify a zeta nonvanishing theorem without covering every Hecke L-function claim or downstream application in the paper105.

Why the Math Community Is Furious

The anger is less about the results than the process. OpenAI bypassed peer review entirely, releasing raw manuscripts to GitHub — even the announcement's own framing acknowledges disregard for long-standing academic etiquette1. Mathematicians were described as furious, caught between awe at the substance and offense at the delivery1.

Verification is also materially harder because neither the model nor the prompts are public. Without model access, outside experts cannot reproduce the generation process, only audit the outputs — a fundamentally different epistemic posture than mathematics is used to8. Will Depue, who built citedbyagi.com to track which human papers the release cites, expects some results will not survive scrutiny2. Levent Alpöge called the quasi-Riemann and no-Siegel-zeros results "the most significant moment in mathematical history," while also flagging scooping and conflict-of-interest problems involving other labs' users2.

The backdrop compounds the unease. WIRED reported that OpenAI privately convened roughly 40 mathematicians in August to discuss what happens if AI surpasses human mathematicians, with attendees saying OpenAI indicated its model had already solved hundreds of difficult open problems5. Anthropic has since disclosed Riemann-adjacent progress of its own, turning mathematics into the newest front of the AI-lab competition5. And OpenAI consulted the Institute for Advanced Study's independent Advisory Group on Mathematics and AI on how to release the work — which lends process legitimacy, but also confirms the release was choreographed in advance while the community at large was kept in the dark24.

The AI-for-Science Stakes

The release matters far beyond number theory. If roughly three hours of frontier-model reasoning can crack problems that have resisted the best human minds for decades2, the bottleneck in theoretical science shifts from generating ideas to verifying them — and formal tools like Lean are the scalable verification substrate. The manuscripts reportedly exhibit "intuitive transfer," constructive counterexamples, and physical intuition drawn from heat-flow simulation and Hamiltonian systems — hallmarks of research-level reasoning rather than retrieval1. Su Weijie, the OpenAI researcher and COPSS Presidents' Award winner, called it "the beginning of a Copernican paradigm shift in humanity's understanding of intelligence"1.

My reading: the quasi-Riemann claim is plausible precisely because it is framed correctly — a 7/8 zero-free region with Lean artifacts, not the full Riemann hypothesis — and because it landed with falsifiable, machine-checkable certificates rather than mere assertion. The Landau–Siegel zero exclusion is arguably the more consequential claim for analytic number theory, since it removes the obstacle that has blocked uniform prime-distribution results for a century16. A polylogarithmic bound on the least quadratic nonresidue, with a deterministic polynomial-time square-root algorithm mod p, follows as a concrete algorithmic consequence10.

But the fury is warranted too. Mathematics is a communal verification process, and a single lab unilaterally dumping 722 unrefereed claims — generated by a model it will not share — inverts the discipline's norms overnight. The next months will be a stress test: if the Lean certificates reproduce and number theorists digest the 199-page zeta argument, the release becomes a milestone in AI-for-science. If they crumble, it becomes a cautionary tale about scale outrunning verification. Either way, the era in which frontier AI models are participants in research mathematics, not just tools, has formally begun21.