OpenAI's AI Disproved an 80-Year Erdős Conjecture (2026)

TL;DR

On May 20, 2026, OpenAI announced that one of its general-purpose reasoning models autonomously disproved the Erdős unit distance conjecture, an 80-year-old open problem in geometry. It found a new family of point arrangements that beats the square grid mathematicians long believed was optimal.¹ The model was not a math-specialized system — it was an unreleased, experimental reasoning model that worked from a single open-ended prompt and produced a 125-page chain of thought.² Nine independent mathematicians, including Fields Medalist Tim Gowers, verified the proof and published a companion paper.³ OpenAI's mathematician Sébastien Bubeck said he believes it is the first time AI has autonomously produced an important result in any field of research.² The result is real — but humans still cleaned it up, and a Princeton mathematician published a sharper version the same day.⁴⁵

What You'll Learn

• What the Erdős unit distance conjecture is, in plain terms
• Exactly what OpenAI's AI proved — and what it did not
• How the AI disproved the conjecture using algebraic number theory
• Which AI model did it, and why OpenAI won't name it
• Why OpenAI's October 2025 Erdős claim collapsed, and how this is different
• Whether this is really the first AI-solved math problem
• What the result means for AI in scientific research

What Is the Erdős Unit Distance Conjecture?

The unit distance problem is one of those rare math questions that a child can understand but no expert could crack. Draw some dots on a sheet of paper. The goal: make as many pairs of dots as possible sit exactly one unit apart — say, one inch. For any number of dots, what is the largest number of unit-distance pairs you can force?⁴

Small cases are easy. Put nine dots in a straight line and you get eight pairs an inch apart. Arrange the same nine dots in a 3×3 square grid and you get twelve.⁴ The grid wins. The hard question is what happens as the number of dots grows into the millions, billions, or beyond.

In 1946, the Hungarian mathematician Paul Erdős (1913–1996) studied the problem and proposed an answer.² He showed that a grid with very carefully chosen spacing — fine enough that unit distances appear across many grid steps, not just neighbors — produces a count that grows just barely faster than the number of dots itself. Then he made a bolder claim: nobody could ever do meaningfully better. That claim became known as the unit distance conjecture, and for eight decades it held. Most mathematicians believed Erdős was right; the few who looked for a counterexample found nothing.⁴

What OpenAI's AI Actually Proved

On May 20, 2026, OpenAI said its AI had broken the record. "For nearly 80 years, mathematicians believed the best possible solutions looked roughly like square grids," the company wrote. "An OpenAI model has now disproved that belief, discovering an entirely new family of constructions that performs better."⁶

The precise achievement is a disproof. Erdős conjectured that no arrangement could beat his grid by more than a vanishingly small margin. The AI produced an arrangement that beats it by a genuine polynomial margin — the kind of improvement Erdős's conjecture said was impossible.¹⁴ In other words, the AI did not just nudge the record; it showed the 80-year-old belief was simply wrong.

Two limits are worth stating clearly. First, the AI did not find the optimal arrangement — it proved a better one exists, but the true maximum for the unit distance problem is still unknown.⁴ Second, the AI's construction was not the last word even on the lower bound: the same day OpenAI announced the result, Princeton mathematician Will Sawin published a follow-up paper, "An explicit lower bound for the unit distance problem," showing arrangements with more than n^1.014 unit-distance pairs — an explicit, sharper version of what the model had produced.⁵

How the AI Disproved the Conjecture

The surprising part is where the proof came from. Generations of geometers had attacked the unit distance problem with the tools of discrete geometry and combinatorics. The AI reached instead for algebraic number theory — a branch of math with no obvious connection to dots on a page.²

Rather than arranging points on a flat sheet directly, the model chose points whose coordinates are solutions to particular algebraic equations.² It built a high-dimensional lattice with special internal symmetries, then mapped that structure back down to two dimensions — a flattened numerical "shadow" of the higher-dimensional object. The result is so far from a simple grid that, mathematicians say, it is too intricate to actually draw on paper, even for a small number of dots.⁴ To prove that the exotic number systems it needed actually exist, the model used heavy machinery: infinite class field towers and the Golod–Shafarevich theorem.⁴

None of these tools is new. What was new was connecting them to the unit distance problem. As Bubeck put it: "The model did not invent something fundamentally new that nobody saw coming. It just executed like an amazing mathematician."⁴ Harvard's Melanie Matchett Wood offered a humbling read: she believes that if the experts who later parsed the AI's answer had instead spent that effort hunting for a counterexample, they probably would have found one themselves.⁴

Which AI Model Solved It?

OpenAI has not named the model. The company says only that it was an experimental, general-purpose reasoning model — explicitly not a system designed for mathematics, and not the kind of specialized prover like Google DeepMind's AlphaProof.²

Bubeck stressed how hands-off the process was: the model received a single prompt — a machine-rewritten statement of Erdős's question asking, open-endedly, whether the conjecture was true or false — and produced one continuous, very long chain of reasoning.² That reasoning runs to a 125-page document, which OpenAI has not fully released.² The company has also kept the model's identity and exact training details private, which is part of why coverage has been careful to call this an OpenAI claim backed by independent verification, rather than a fully open result.⁶²

If you have followed OpenAI's recent reasoning-model push, the framing fits a pattern: the company keeps arguing that general reasoning, not task-specific tuning, is what unlocks hard problems.

The October 2025 Erdős Problem Embarrassment

This is not OpenAI's first Erdős headline — and the first one went badly. Around October 2025, OpenAI's then-VP Kevin Weil, who led the company's OpenAI for Science team, posted on X that "GPT-5 found solutions to 10 (!) previously unsolved Erdős problems and made progress on 11 others."⁷

It wasn't true. GPT-5 had not solved anything new; it had run a literature search and surfaced existing published solutions that the community simply hadn't catalogued.⁷ Thomas Bloom — the mathematician who maintains the erdosproblems.com database, where "open" means only that he personally knew of no solution — called the post a dramatic misrepresentation. Weil deleted it, and rivals including Yann LeCun and Google DeepMind's Demis Hassabis piled on.⁷

The May 2026 announcement was built to avoid a repeat. The unit distance result is an original disproof, not a literature lookup. And the same Thomas Bloom who exposed the 2025 misfire is now a signatory on the paper verifying the new one — a deliberate signal that this time the math holds up.⁶³

Is This Really the First AI Math Breakthrough?

OpenAI's exact wording is careful: it calls this "the first time AI has autonomously solved a prominent open problem central to a field of mathematics."⁶ Each qualifier matters, because AI has touched serious math before.

Google DeepMind's AI reached silver-medal standard at the 2024 International Mathematical Olympiad, led by its AlphaProof system, and a Gemini Deep Think system hit gold-medal level in 2025 — but Olympiad problems are competition puzzles with known answers, not open research questions.⁸ DeepMind's AlphaEvolve did tackle open problems in 2025, even improving the known lower bound for the kissing number in 11 dimensions from 592 to 593 — but AlphaEvolve is an evolutionary system scaffolded to search through candidate solutions.⁸ And in January 2026, GPT-5.2 generated proofs for three genuinely open Erdős problems — formalized in the Lean proof assistant and verified by Fields Medalist Terence Tao, who framed them as the easier "long tail" of Erdős problems rather than deep breakthroughs.⁹

What makes the unit distance result stand out is the combination: a well-known, decades-old, much-attempted problem, solved by a general-purpose model, autonomously, from one prompt. That is why Daniel Litt of the University of Toronto — one of the verifiers — called it "the first result produced autonomously by an AI that I find interesting in itself."² The honest summary: it is a milestone, and the superlatives belong to OpenAI and the mathematicians who checked the work, not to a marketing line.

What It Means for AI and Mathematics

The verification process is the real story for anyone tracking AI's research credibility. OpenAI privately sent the proof to independent mathematicians, who — without the company's involvement — wrote a 19-page companion paper, "Remarks on the disproof of the unit distance conjecture," signed by nine researchers including Noga Alon, Thomas Bloom, Tim Gowers, Daniel Litt, Will Sawin, Jacob Tsimerman and Melanie Matchett Wood.³ Gowers, a Fields Medalist, said no previous AI-generated proof had come close to the standard of a top journal.⁴

But the caveats are just as important. No outside expert saw the model's raw output — only an edited version of its reasoning.⁴ And the result needed human hands to become a real proof. As Bloom put it, "The human still plays a vital role in discussing, digesting, and improving this proof."⁴ This was a collaboration, not a handoff.

The forward-looking claim from OpenAI is that models able to hold together long, cross-disciplinary chains of reasoning could help in biology, physics, engineering and medicine — the same bet behind efforts to point models at scientific discovery and drug research.⁶ Whether that generalizes is unproven. For now, the lesson echoes what we saw when AI models began clearing hard agentic benchmarks: a real capability jump, wrapped in claims worth reading slowly.

The Bottom Line

OpenAI's disproof of the Erdős unit distance conjecture is the most credible AI-driven math result so far, and the verification by nine mathematicians — Tim Gowers among them — is what separates it from the company's embarrassing 2025 misfire. A general-purpose reasoning model, working from one prompt, connected algebraic number theory to a geometry problem that had resisted experts for 80 years. That is genuinely new.

It is also worth keeping in proportion. The model did not find the optimal answer, its construction was sharpened by a human mathematician within a day, and the proof only became publishable after significant human cleanup. The accurate way to describe what happened is not "AI replaced mathematicians" but "AI became a startlingly capable collaborator." For developers and researchers watching where reasoning models are headed, that distinction — capability paired with claims you should still read carefully — is the whole story.