On May 20, 2026, OpenAI announced that an internal, general-purpose reasoning model had autonomously disproved the planar unit distance conjecture Paul Erdős posed in 1946 — one of the central open problems of discrete geometry. The result did not come from a math-specialized model wrapped in search scaffolding: the proof emerged from a general model, and along the way it chose to hunt for a counterexample rather than try to prove the bound everyone believed.
Timothy Gowers, the Fields medalist, called it “a milestone in AI mathematics” and said that submitted to the Annals of Mathematics, he would have recommended acceptance “without any hesitation.” An eighty-year-old question, closed by a model that was never built for mathematics.
An Eight-Decade-Old Question Falls
The unit distance problem asks: among n points in the plane, how many pairs can sit at distance exactly 1? Erdős conjectured in 1946 that the answer is n^(1+o(1)) — essentially linear. For nearly eight decades the best lower bound remained, in effect, Erdős’s own rescaled integer grid, on the order of n^(1+C/log log n), while the upper bound froze at O(n^(4/3)) from Spencer, Szemerédi, and Trotter in 1984. Nobody had moved the gap.
The model produced a third answer: for infinitely many n, there exist point sets with at least n^(1+δ) unit-distance pairs — which kills the conjecture outright. Will Sawin at Princeton then posted a paper to arXiv on May 20 (2605.20579) making the exponent explicit: n^1.014.
How the Proof Works: Number Theory Enters
The most surprising part is where the tools came from. Erdős’s original construction uses Gaussian integers; the model’s breakthrough is to swap in richer algebraic number fields whose symmetries produce more differences of length exactly 1. The existence of suitable fields follows from infinite class field towers and Golod–Shafarevich theory — machinery from deep algebraic number theory, showing up in a problem that looks like elementary Euclidean geometry.
Sawin’s paper follows the same route, using a Golod–Shafarevich criterion argument to build fields of large degree and small discriminant containing many primes of small norm, pushing δ to 0.014. Thomas Bloom’s remarks note that the result shows number-theoretic constructions have far more to say about discrete geometry than anyone suspected.
Verification and the Mathematicians’ Response
OpenAI shipped supporting material alongside the announcement: a companion explanatory paper, “Remarks on the Disproof of the Unit Distance Conjecture,” written by external mathematicians, plus an abridged chain of thought. Noga Alon of Princeton said “every combinatorial geometer” had thought about the problem and called the algebraic number theory tools “elegant and clever.”
Arul Shankar observed that the model’s chain of thought mostly attempted counterexamples rather than proving the believed bound, concluding that AI models “are capable of having original ingenious ideas.” Jacob Tsimerman admitted he had failed at a similar counterexample attempt himself. The O(n^(4/3)) upper bound still stands, so the real question shifts: what is the true growth rate?
Why It Matters
Three layers. First, this is the first autonomous AI solution of a prominent open problem central to a mathematical subfield — and it came from a general-purpose model, meaning the capability lives in the reasoning, not in domain scaffolding. Second, it opens an unexpected bridge between algebraic number theory and discrete geometry, giving mathematicians genuinely new territory. Third, for the trajectory of AI research itself: OpenAI’s own framing is that models can act as research partners across the sciences while humans still choose the problems and interpret the results — and that progress of this kind makes alignment more urgent, not less.
One caveat tempers the announcement: the model is internal and unnamed, so outside teams can neither run it nor compare it against their own stacks today. What is independently checkable is the mathematics — the disproof and Sawin’s refinement now exist as papers any reader can verify line by line, which is exactly why mathematicians took the result seriously rather than treating it as a vendor demo.
It points the same way as Google’s ERA agent writing expert-level scientific code: models are starting to leave verifiable output in research itself, not just help with the writing. For teams building research tools, the signal is clear — the next generation of products should assume models can make original research contributions.
Sources
- An OpenAI model has disproved a central conjecture in discrete geometry — OpenAI
- An explicit lower bound for the unit distance problem — Will Sawin, arXiv
- Remarks on the Disproof of the Unit Distance Conjecture — OpenAI (PDF)
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.