OpenAI cracked an 80-year math problem without a math-specialized model

OpenAI’s general-purpose reasoner just broke an 80-year-old conjecture in discrete geometry. The model wasn’t trained to do mathematics. It runs the same architecture that drafts emails and writes Python, and on Tuesday it produced a new family of geometric configurations that four mathematicians have now verified.

The problem in question has a deceptively simple statement. Take n points in a plane. How many pairs of them can sit at exactly the same distance from each other, say one unit apart? Paul Erdős posed the question in 1946 and offered an upper bound: roughly n to the (1 plus o(1)), shorthand for “barely more than linear.” For decades the best-known configurations came from variants of the square grid, and the grid sat very close to that ceiling. Working mathematicians treated the bound as essentially tight.

OpenAI’s model didn’t tighten the bound. It broke it. The system produced an entire family of point arrangements with at least n to the (1 plus δ) unit-distance pairs, for some fixed δ greater than zero. That isn’t a refinement; it’s a counterexample to the conjecture’s central claim. Will Sawin, one of the four mathematicians who reviewed the work, refined the new exponent into a clean expression. Thomas Bloom, Melanie Wood and Noga Alon, the others on the verifying team, confirmed the construction held.

You May Also LikeHow Gemma 4 and DeepSeek V4 ended the brute-force era of LLM scaling

What’s interesting about the method is that it didn’t come from inside geometry. The model crossed into algebraic number theory, extending the Gaussian integers into other algebraic number fields and treating the resulting lattice points as candidate configurations. That bridge, geometry pulled into number theory, was the move humans had missed for eight decades. It is the kind of jump that, in a math seminar, gets a slow nod and a long silence.

Reactions from working mathematicians arrived inside the first day. Timothy Gowers, who holds a Fields Medal, called it “the first really clear example of AI solving a really well-known math problem.” Alexander Wei, an OpenAI researcher, wrote that the result is the kind a referee at the Annals of Mathematics would accept “without any hesitation.” That last claim is testable. The proof has been published as a PDF with a companion remarks document, and the broader math community is now reading.

The framing OpenAI is leaning on is that this is the first time an AI system has autonomously solved a prominent open problem central to a field of mathematics. The verb “autonomously” is doing a lot of work there. The model produced the construction; the proof was vetted, refined and stress-tested by four human mathematicians before any announcement went out. That distinction matters, because OpenAI has been here before.

In October 2025 the company circulated a claim that a different internal model had solved ten open problems posed by Erdős. Within days, mathematicians showed that several of the “solutions” were either already known or simply wrong. OpenAI retracted the broader claim. That episode is the reason this week’s announcement leads with the names of the verifiers rather than the name of the model. The four mathematicians are the warranty.

The other detail worth holding onto is what kind of model produced the result. OpenAI hasn’t disclosed the system’s name, only that it is a general-purpose reasoning model, the same family of systems that handles chat, drafts code and answers customer-service tickets. There is no math-specialized variant in the loop. The same architecture that handles everyday conversation handled this. The implication is that the bottleneck for AI-driven mathematics may not have been a math-tuned model. It may have been compute and patience.

That bottleneck breaking is the actual story. For a long time the working assumption among researchers was that genuinely original mathematics would require purpose-built systems: theorem provers, formal-verification frameworks, narrow models trained on a corpus of proofs. What landed on Tuesday is a different kind of evidence. A reasoner pointed at a famous, unsolved, eighty-year-old problem; given enough room to think, it produced something Sawin, Bloom, Wood and Alon agreed was correct. The path from chat windows to Erdős turned out to be shorter than expected.

Recommended ReadingAnthropic bought Stainless and is shutting down the SDK pipeline OpenAI used

Some caveats still apply. The model is not publicly available. Independent groups outside the original four-mathematician panel will read the proof in the coming weeks, and the broader peer-review process for the Annals or another top journal will take months. The exponent δ is small. The construction does not solve the related unit-distance problem on the sphere or in higher dimensions. None of that diminishes what happened on Tuesday. It places it.

What changes is the expectation. A year ago, the question for AI in mathematics was whether systems could ever produce original proofs of consequence. As of this week, the question is which open problem falls next, and whether the mathematicians who verify the proofs will keep being credited the way Alon and his colleagues were here.

A 1946 conjecture is one of those quiet objects that sits on a shelf, waiting for the right hand to take it down. The hand that took it down this week ran on a GPU cluster, had not been trained for the job, and finished the work while four mathematicians watched.

More Like This

Two points behind Opus 4.6, five times cheaper: Gemini 3.5 Flash redraws the math