The AI-assisted math results were first coming in a trickle, but they seem to be threatening to turn into a deluge.
OpenAI has announced that one of its internal reasoning models has solved the planar unit distance problem, a foundational open question in combinatorial geometry first posed by the prolific Hungarian mathematician Paul Erdős in 1946. The result has been independently verified by a group of leading external mathematicians, and the proof has been called “a milestone in AI mathematics” by Fields Medalist Tim Gowers.
What Is The Unit Distance Problem?
The question, deceptively simple to state, goes like this: if you place n points in a plane, what is the maximum number of pairs of points that can be exactly distance 1 apart? Erdős, who spent his career collecting and distributing unsolved problems across mathematics, even attached a monetary prize to this one.
For nearly 80 years, the prevailing belief was that the answer looked roughly like a rescaled square grid. The best known construction, based on the so-called Gaussian integers, gives approximately n^(1+C/log log n) unit-distance pairs — only slightly faster than linear growth. Erdős conjectured that no construction could do fundamentally better, and the mathematical community largely agreed.
The OpenAI model has now proved that conjecture wrong.
The Breakthrough
The model produced a proof showing that for infinitely many values of n, you can arrange n points to yield at least n^(1+δ) unit-distance pairs, for some fixed positive exponent δ. That is a polynomial improvement — a qualitatively different order of magnitude — over what anyone previously thought possible. A refinement by Princeton professor Will Sawin subsequently pinned down δ = 0.014.
What makes the result especially striking is how it was found. The proof doesn’t use better geometry. It imports ideas from algebraic number theory — specifically, tools such as infinite class field towers and Golod–Shafarevich theory, concepts developed to study factorization in extensions of the integers — and turns them onto an elementary-sounding question about distances in the Euclidean plane. The connection was entirely unexpected.
Tim Gowers put it plainly: “There is no doubt that the solution to the unit-distance problem is a milestone in AI mathematics: if a human had written the paper and submitted it to the Annals of Mathematics and I had been asked for a quick opinion, I would have recommended acceptance without any hesitation.”
Part Of A Broader Pattern
This is not a one-off. AI models have been racking up mathematical milestones at an accelerating pace. OpenAI’s models have won gold medals at the International Mathematical Olympiad. GPT-5.2 helped solve Erdős Problem #281, which Fields Medalist Terence Tao called “perhaps the most unambiguous instance” of AI solving an open mathematical problem. Harmonic, the math AI startup backed by Robinhood CEO Vlad Tenev, claimed to have cracked Erdős Problem #124, open for 30 years. And even Donald Knuth, one of computing’s founding giants and a longtime AI skeptic, was recently left stunned when Claude solved an open problem he had been working on for weeks.
That said, the field has also seen overclaiming. Google DeepMind CEO Demis Hassabis publicly called out OpenAI earlier for suggesting GPT-5 had “solved” Erdős problems when the model had actually retrieved known solutions from the internet. The unit distance result is in a different category: it is a genuinely new proof, verified by independent experts, resolving a conjecture that had stood essentially unchanged since 1946.
Why This One Is Different
Most previous AI math achievements — however impressive — involved either competition-style problems with known solution structures, or finding existing proofs the model had been trained near. This result involved autonomous discovery: the model was given a collection of Erdős problems as a test, with no special scaffolding for the unit distance problem in particular, and produced a proof that resolves a central open conjecture.
The proof also required genuine cross-domain creativity. No one working in discrete geometry had thought to apply Golod–Shafarevich theory to counting unit distances. The model apparently did. Thomas Bloom, who curates the Erdős problems database and contributed to the companion paper, noted that the result “shows that there is a lot more that number theoretic constructions have to say about these sorts of questions than we suspected.”
What Comes Next
OpenAI is clear that this is a signal, not just a result. The same capabilities that enable a model to hold together a long mathematical argument — coherent multi-step reasoning, connecting disparate fields, generating novel structures — are the capabilities that matter in biology, physics, materials science, and drug discovery.
Mathematics is, in some ways, an ideal proving ground: the problems are precise, the proofs are checkable, and there is no ambiguity about whether the answer is right. That the bar has now been cleared at this level — a prominent open problem, central to an active subfield, solved autonomously — suggests that the next phase of AI-assisted research is arriving faster than most expected.