Summary: On May 20, 2026, OpenAI said an internal general-purpose reasoning model disproved Paul Erdős’s 1946 planar unit-distance conjecture without mathematics-specific tools or search scaffolding. External mathematicians reviewed supporting material, and Will Sawin posted a same-day refinement giving an n^1.014 lower bound while the O(n^(4/3)) upper bound still stands. The mathematics is independently checkable, but the unnamed internal model cannot be tested by outside teams.
The news broke on May 20, 2026: an internal general-purpose reasoning model, tested at OpenAI, had on its own knocked down the planar unit distance conjecture Paul Erdős posed in 1946 — one of the central unsolved problems of discrete geometry. No math-specialized tools or search scaffolding produced the result: the proof came from a general model with no special mathematical training, which opted to hunt a counterexample rather than establish the bound everyone believed.
Fields medalist Timothy Gowers branded the result “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 deadlock, closed out by a model never built for mathematics.
A Question That Stood for Eighty Years Comes Undone
The problem takes one sentence to state: scatter n points across the plane — at most how many pairs can sit exactly 1 apart? Erdős guessed in 1946 that the answer runs n^(1+o(1)) — close to linear. For the next eight decades the only lower-bound construction that counted stayed Erdős’s own rescaled integer grid, near n^(1+C/log log n), and the upper side never moved past the O(n^(4/3)) Spencer, Szemerédi, and Trotter proved in 1984. The gap defeated everyone who attacked it.
The model delivered a third kind of answer: for infinitely many n, point sets exist carrying at least n^(1+δ) unit-distance pairs — the conjecture falls outright. Will Sawin of Princeton followed the same day with an arXiv posting (May 20, 2605.20579) pinning the exponent at n^1.014.
Where the Proof Came From: Algebraic Number Theory Takes the Stage
The technical shock is where the tools came from. Erdős built his construction from Gaussian integers; the model’s breakthrough swaps in richer algebraic number fields, whose added symmetries manufacture more differences of length exactly 1. That such fields exist rests on infinite class field towers and Golod–Shafarevich theory — deep algebraic number theory machinery, turning up in a problem that looks like elementary Euclidean geometry.
Sawin’s paper takes the same road: a Golod–Shafarevich criterion argument builds fields of large degree and small discriminant, rich in primes of small norm, lifting δ to 0.014. Thomas Bloom’s remarks note that the result shows number-theoretic constructions have far more to offer discrete geometry than anyone suspected.
How Mathematicians Checked It, and What They Said
Alongside the announcement came supporting material: an 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” has thought about the problem and judged the algebraic number theory tools “elegant and clever.”
Arul Shankar noticed the model’s chain of thought mostly chased counterexamples rather than proving the conjecture, concluding that AI models “are capable of having original ingenious ideas.” Jacob Tsimerman owned up to failing on a similar counterexample attempt himself. With the O(n^(4/3)) upper bound still standing, the live question moves: what is the correct growth rate?
What Makes This a Big Deal
Three layers of significance. First, an AI system has, unaided, solved a prominent open problem central to a mathematical subfield for the first time — and a general-purpose model did it, so the capability lives in the reasoning, not in domain scaffolding. Second, an unexpected bridge now links algebraic number theory with discrete geometry, handing mathematicians genuinely new problems. Third, OpenAI itself frames models as research partners across the sciences while people keep picking the problems and reading the results — and treats progress of this kind as making alignment more urgent, not less.
One caveat shadows the announcement: the model is internal and goes by no public name, so outside teams cannot run it or test it against their own stacks today. What anyone can independently check is the mathematics — the disproof and Sawin’s refinement exist as papers verifiable line by line, which is why mathematicians engaged with the result rather than filing it as a vendor demo.
The trajectory matches Google’s ERA agent writing expert-level scientific code in Nature: models have begun leaving verifiable output inside the research itself rather than the prose around it. For research-tool teams the reading is clear — build the next product generation on the assumption that models can make original research contributions.