An unnamed internal model claims to have cracked one of the seven Millennium Prize Problems — and hundreds more besides. The math world’s response? Show your work, or stop.
On the morning of October 6, 2026, a GitHub repository appeared without warning. No blog post. No press release. No named author. Just a folder called `openai/math`, containing 722 manuscripts, 372 result families, and a set of claims that — if even a fraction are correct — would rank among the most significant events in the history of mathematics.
By afternoon, mathematicians around the world were reading them. By evening, the field was in revolt.
A Model That Doesn’t Have a Name
The manuscripts are credited to an internal OpenAI model that the company has declined to name, describe, or release for outside testing. What is known: training reportedly began on August 28, 2026. By September 21 — just 24 days in — OpenAI announced that the model had already produced progress on “more than 100” long-standing open problems across number theory, combinatorics, theoretical computer science, and mathematical physics. The model was burning through roughly three hours of ChatGPT Pro-equivalent compute per accepted result, averaging problems that human mathematicians had puzzled over for decades.
The October 6 release expanded the catalog to 722 manuscripts. By OpenAI’s count, the model had engaged with roughly 4,000 unsolved problems — a number so large it essentially represents the research agenda of an entire generation of professional mathematicians, attempted in a matter of weeks.
The most dramatic claim involves the Navier-Stokes existence and smoothness problem — one of the seven Millennium Prize Problems identified by the Clay Mathematics Institute in 2000, each carrying a $1 million prize and representing an unsolved frontier of human knowledge. OpenAI says the model produced a proof of singularity in the 3D Navier-Stokes equations, accomplished using a swarm of approximately 10,000 coordinating agents over roughly 88 hours and nearly 5 million messages.
To understand why this matters: the Navier-Stokes equations govern fluid dynamics — how air flows over a wing, how blood moves through arteries, how ocean currents evolve. Whether solutions to these equations can break down (form singularities) is one of the deepest open questions in applied mathematics. No human mathematician had cracked it. OpenAI’s unnamed AI reportedly did it in under four days.
The Asterisks Are Everywhere
Within 24 hours, the mathematical community had found the first problem.
On October 7, OpenAI quietly withdrew three of the 722 manuscripts. The reason: a sign error in one paper invalidated its argument, and two dependent papers collapsed with it. Over the following week, 14 more manuscripts were revised. As of mid-October, roughly 50% of the released results had not been fully verified before publication, according to OpenAI’s own disclosure to Retraction Watch. Of the 719 remaining manuscripts, only about 42% had been put through Lean — the formal verification system mathematicians use to check proofs with machine precision.
Princeton mathematician Charles Fefferman — one of the world’s foremost experts on fluid dynamics — noted that credit for the key insight behind the Navier-Stokes result likely belongs to human mathematicians Diego Córdoba and Luis Martínez-Zoroa, whose prior work the AI appears to have built upon. NYU mathematician Tristan Buckmaster, who had seen an early AI-generated draft of the proof, described it as “badly written and hard to parse.” NPR reported that the research community had learned “comparatively little” from the result because OpenAI had not published the model’s reasoning in any form others could study or replicate.
Only 10 of the 722 manuscripts included even abridged reasoning summaries. The model’s prompts were not published. The model itself remains unavailable. What OpenAI released, critics argued, was not mathematics — it was the output of mathematics, stripped of the context that gives mathematical work its meaning.
Twenty-Five Fields Medalists Sign a Letter
The Fields Medal is mathematics’ highest honor, typically awarded once every four years to researchers under 40 for contributions “of the highest significance.” Getting 25 of them to agree on anything is unusual. Getting 25 of them to sign a joint open letter is extraordinary.
The letter, which predates the October 6 release, warned that AI’s growing role in mathematical research could undermine the field’s standards and its fundamental purpose. The mathematicians did not dispute that AI systems could generate proofs. They disputed whether generating proofs at scale, without human understanding, constituted mathematical progress at all.
The Association for Human Mathematics, writing on Fields Medalist Terence Tao’s blog shortly after the release, was more pointed. The group called the dump of 700+ files “a demonstration of power rather than scholarship,” argued that OpenAI had violated the explicit recommendations of the Advisory Group on Mathematics and AI (which had urged labs to stop testing advanced problems on proprietary models), and called on mathematicians to discontinue working with OpenAI entirely. “We reject OpenAI’s assertion that this release advances our subject,” the statement read. “Mathematicians did not ask for this work to be done.”
The verification burden, critics noted, now falls entirely on the research community. OpenAI has effectively uploaded hundreds of research leads and told the world’s mathematicians to figure out which ones are real — without providing the model, the methodology, or in most cases the reasoning that generated them.
What OpenAI Says
OpenAI has acknowledged the withdrawals and revisions but maintained that the release represents genuine progress. The company reportedly formed an advisory group, AGMAI, and consulted the Institute for Advanced Study before publishing, though it is unclear whether that group reviewed all 722 manuscripts. OpenAI has said it expects outside experts to find errors and request revisions — framing the GitHub repository less like a collection of finished results and more like a living document, subject to ongoing community correction.
That framing has not satisfied critics. The Advisory Group on Mathematics and AI, after reviewing more than 600 community responses, published recommendations in late September warning against exactly this kind of release: labs should not test frontier models on advanced problems and publicize selective results. Measured against the group’s checklist, OpenAI’s release is only partially compliant. The repository is hosted on OpenAI’s own GitHub account. The model is unnamed. The prompts that generated the results are not published.
The Deeper Question
Behind the technical debates lies a harder question that the mathematical community has been grappling with since AI first began to show research-level capability: what is mathematics for?
If a machine can generate 722 proofs in six weeks that would take thousands of human-years to produce independently, does it matter that humans cannot understand the machine’s reasoning? The history of mathematics is not simply a catalog of true statements — it is a record of understanding. When humans proved the Four Color Theorem using computers in 1976, the math community spent decades arguing about whether a computer-assisted proof counted. The OpenAI release scales that argument by orders of magnitude, and removes the human mathematicians from the loop almost entirely.
The Millennium Prize Problems were designed to represent the frontier of what human minds could not yet grasp. OpenAI’s unnamed model may have crossed that frontier. Whether it has helped anyone understand what lies on the other side is a different question — and for many mathematicians, it is the only one that matters.
For now, the 719 manuscripts sit on GitHub, partially verified, heavily disputed, and carrying the weight of some of the most audacious mathematical claims in living memory. The math community is reading. It will take a while.
Sources: AIWeekly, Tech-Insider, Nature, Terence Tao’s blog (AHM Statement), Quanta Magazine, Retraction Watch, NPR
