OpenAI has made this kind of announcement before, and it did not go well. Seven months ago, a company executive posted on X that GPT-5 had solved ten previously unsolved Erdős problems. Within days, it emerged that GPT-5 hadn’t solved them at all — it had found solutions that already existed in the mathematical literature, solutions the model apparently didn’t know it was retrieving. The post came down. Yann LeCun posted. Demis Hassabis posted. The internet dunked, correctly.
So when OpenAI announced today that a reasoning model had autonomously disproved a conjecture first posed by Paul Erdős in 1946, the appropriate first response was skepticism. The appropriate second response, after reading what actually happened, is to update significantly. This one appears to be real — and the way it was done is as interesting as the result itself.
- The problem
- The planar unit distance problem — first posed by Erdős in 1946. Among n points in a plane, what is the maximum number of pairs at exactly distance 1 apart?
- The conjecture
- That square-grid constructions are essentially optimal: growth rate at most n1+o(1), where the additional term vanishes as n grows
- The result
- An OpenAI reasoning model disproved this, constructing configurations with at least n1.014 unit-distance pairs — a genuine polynomial improvement
- Verified by
- Tim Gowers (Fields Medal), Noga Alon (Princeton), Melanie Wood, Thomas Bloom, Will Sawin (Princeton, refined the exponent to δ = 0.014)
- What it used
- Algebraic number theory — specifically infinite class field towers and Golod-Shafarevich theory — applied to a geometry problem for the first time
The Problem, Explained Without a Math Degree
Put some dots on a piece of paper. Now count how many pairs of dots are exactly one unit apart. The question Erdős asked in 1946 was: if you place n dots as cleverly as possible, what’s the maximum number of pairs you can get at distance 1?
The naive answer — a line of dots, equally spaced — gives you n−1 pairs. A square grid does better: roughly 2n pairs. And if you scale that grid carefully, you can squeeze out a bit more. The best known construction before today achieved n raised to the power of 1 plus C divided by log(log(n)) — which sounds impressive but is barely faster than linear, because log(log(n)) grows so slowly that the extra term in the exponent is nearly zero for any n you’d actually work with.
Erdős conjectured this was essentially the ceiling. The prevailing belief for eight decades was that no construction could do significantly better than the square grid. Today, that belief was disproved.
What the AI Actually Found
The new proof constructs configurations of n points with at least n1+δ unit-distance pairs, for some fixed positive exponent δ. Princeton mathematician Will Sawin subsequently refined the argument to show you can take δ = 0.014. That’s a polynomial improvement over the square grid — not a marginal gain, but a qualitatively different rate of growth that holds for infinitely many values of n.
The surprising part isn’t the result. It’s where the proof came from. The tools used — infinite class field towers, Golod-Shafarevich theory — are well-established ideas from algebraic number theory, a branch of mathematics that studies how integers behave in extended number systems. Nobody had applied them to a question about dots and distances in the Euclidean plane. The connection wasn’t in the literature. It wasn’t in the existing approaches to the problem. An AI system found it independently.
Thomas Bloom, who maintains the Erdős Problems website and who publicly called out OpenAI’s false claim seven months ago, put it plainly in the companion paper: this result shows “there is a lot more that number theoretic constructions have to say about these sorts of questions than we suspected; moreover, that the number theory required can be very deep.”
Why This Time Is Different
The October 2025 embarrassment happened because someone at OpenAI announced a result before it had been checked against the literature. The model found things that looked like novel solutions and weren’t. This time, OpenAI did the thing they should have done before: they got mathematicians to read it first.
The model used to produce this result is described as a general-purpose reasoning model — not a system fine-tuned on mathematical literature, not a proof assistant, not scaffolded specifically to search for solutions to this problem. OpenAI says it was evaluating the model on a collection of Erdős problems as part of broader frontier research testing. On this one, it produced a proof.
Fields medalist Tim Gowers, in the companion paper, called it “a milestone in AI mathematics.” Arul Shankar, a leading number theorist, went further: “In my opinion this paper demonstrates that current AI models go beyond just helpers to human mathematicians — they are capable of having original ingenious ideas, and then carrying them out to fruition.”
What This Actually Means
There’s a version of this story that gets overclaimed quickly: AI can now do mathematics, the era of human mathematical discovery is ending. That’s not what happened here, and the mathematicians involved are not saying that.
What happened is more specific and more interesting. A system found a non-obvious connection between two distant areas of mathematics — algebraic number theory and discrete geometry — that human mathematicians had not found in 80 years of working on the problem. It then followed that connection through to a complete proof. That’s a meaningful capability. It suggests these systems can do something more than search through known ideas; they can, at least in some cases, combine them in ways that count as genuinely novel.
The practical ceiling on that capability is still unclear. This is one result, on one problem. The history of AI progress contains many “first times” that turned out to be narrower than they appeared. But the unit distance problem is not an obscure puzzle — it’s been described as “possibly the best known problem in combinatorial geometry.” The people checking the proof are not easily impressed. They’re impressed.
What Happens Next
The immediate effect is that a generation of mathematicians now has a new technique to apply to related problems. The class field tower approach that the AI used to beat the unit distance bound might work on other geometric problems where algebraic number theory hasn’t been tried. Bloom’s observation isn’t a prediction — it’s already happening.
The longer-term implication is harder to assess. OpenAI frames this as evidence that AI systems can contribute to frontier research — not as assistants searching existing literature, but as participants generating novel mathematical ideas. If that’s true at scale, the structure of mathematical research changes. Problems that have resisted decades of effort might yield faster.
Whether this result is a one-off or the beginning of a pattern is genuinely unknown. One result doesn’t prove a capability is reliable. But one result, when it’s this result, checked by these people, found this way — it warrants more than a shrug.
This is a real mathematical result, verified by credible people, that resolves a prominent 80-year-old open problem. The method — applying algebraic number theory tools to a geometry question — was not in the literature. An AI found it autonomously.
OpenAI has cried wolf on AI math before and was wrong. The relevant difference this time is that the mathematicians who caught them last time are the ones saying the proof is correct.
The overclaim to avoid: that this means AI can now do all mathematics. The claim that appears warranted: that AI systems can, at least sometimes, find genuine connections between distant fields that humans haven't found. That's a different kind of capability than we've seen demonstrated before.