On May 20, 2026, OpenAI announced something that made mathematicians stop and stare. An internal reasoning model had independently disproved a conjecture that had stood unsolved for 80 years. Not a competition problem. Not a well-posed theorem that someone had a hunch about. A central open question in discrete geometry, posed by one of the great mathematicians of the 20th century, that professionals had failed to crack across eight decades.

The model didn’t just find the answer. It found a genuinely new one — using a branch of mathematics nobody had applied to this problem before.

Something different is going on here.

The maths, for the rest of us

Imagine you have a handful of dots and a piece of paper. You also have a stick that’s exactly one unit long — say, 10cm.

The question is: if I give you n dots, what’s the maximum number of dot-pairs you can arrange so that the stick fits exactly between them?

Concrete example: place 4 dots as a square, each side exactly 1 unit. You get 4 pairs at unit distance (the four sides). Now place 9 dots in a 3×3 grid, one unit apart. Count the pairs — it’s more, but it grows slower than you might expect as you add more dots.

For 80 years, that square grid was the best arrangement anyone could find. With n dots, the grid gives you roughly n^(4/3) pairs. The conjecture said: that’s about as good as it gets. The grid wins.

What the AI found is that this is wrong. There is a way to place the dots — not in a grid, but using a specific construction from number theory — that produces more pairs at unit distance. A provably, mathematically meaningful amount more.

Here’s where it gets strange. Instead of thinking about geometry, the approach uses algebraic number fields: abstract mathematical structures built from prime numbers and extensions of ordinary arithmetic. It’s as if someone asked “what’s the fastest route from Geneva to Zürich?” and the correct answer turned out to involve a detour through a completely different branch of mathematics that wasn’t on the map.

Nobody had looked there. Not because mathematicians are bad at their jobs. Because human researchers specialise. The number theorists weren’t thinking about dot arrangements. The geometers weren’t fluent in class field towers. The AI had no such boundary.

The problem: dots and distances

The unit distance problem sounds deceptively simple. Place n points anywhere on a flat plane. How many pairs of those points can be exactly one unit apart?

For 80 years, the best answer mathematicians could construct was based on square grid arrangements. Pack your points into a regular grid, and you can achieve roughly n^(4/3) pairs at unit distance. The conjecture — formulated by Paul Erdős in 1946 — was that this was essentially optimal. You couldn’t do meaningfully better.

This matters beyond the counting itself. The unit distance problem sits at the intersection of discrete geometry, combinatorics, and number theory. Progress here tends to open adjacent problems across multiple fields. Which is also why it had stayed unsolved: each field’s standard tools, applied in isolation, weren’t sufficient to crack it. You needed something that crossed boundaries.

The approach nobody tried

The OpenAI model did not iterate on grid arrangements. It went somewhere completely different.

It approached the problem through algebraic number theory — specifically through mathematical structures called Golod-Shafarevich towers, also known as infinite class field towers. These objects belong to a part of mathematics that researchers working on the unit distance problem had simply never reached for.

The key insight was a kind of inversion. Erdős’s original geometric approach was to fix the algebraic number field and vary the prime numbers involved. The model did the opposite: fix the primes, vary the field. That reversal unlocked a construction producing infinitely many point configurations that beat the square grid bound by a polynomial factor.

The result: for infinitely many values of n, you can achieve at least n^(1+δ) unit-distance pairs, where δ is a fixed positive constant. The square grid bound of n^(4/3) is not tight. The conjecture is false.

Will Sawin, a mathematician at Princeton, subsequently examined the proof and tightened the argument, refining δ to a concrete value of 0.014 — making the explicit result n^1.014.

Verification that held up

A 125-page proof from an AI model is not something the mathematical community takes on faith.

Nine external mathematicians checked the work and co-signed a 19-page companion paper: Alon, Bloom, Gowers, Litt, Sawin, Shankar, Tsimerman, Wang, and Matchett Wood. Tim Gowers — a Fields Medalist, the highest honor in mathematics — reviewed the proof and validated its correctness. Sawin went further than validating it; he strengthened it.

One detail worth noting: this was OpenAI’s second attempt on this specific problem. An earlier model had failed. The published result represents a deliberate second run at the same target. That’s relevant context for anyone thinking about how AI-assisted mathematical research actually works — it’s iterative, not a single lucky strike.

What makes this different

Previous AI math results have been genuinely impressive. DeepMind’s AlphaProof verified proofs and solved International Mathematical Olympiad problems at gold-medal level. AlphaGeometry tackled plane geometry problems. These are meaningful milestones.

This is different in kind. Olympiad problems, by design, have known solutions — they are crafted to be solvable within a competition setting. The unit distance problem was an open question at the frontier of active research, without a known answer, where professional mathematicians working in exactly the relevant fields had genuinely failed to make progress for eight decades.

More to the point: the approach was novel. The connection between the unit distance problem and algebraic class field towers was not a known technique that the model retrieved and applied. Human researchers working on this problem had not made this connection. They hadn’t been looking in that direction. The model, working without the same disciplinary specialisation, found a path that wasn’t on anyone’s map.

What it does and doesn’t mean

This doesn’t mean AI can now solve open mathematical problems on demand. The unit distance problem had particular structure — a clean conjecture, a well-defined search space, deep connections across mathematical fields — that played to algebraic reasoning in ways many open problems don’t.

The 125-page proof still required human mathematicians to check, interpret, and strengthen it. The model doesn’t narrate its reasoning in a way that makes the discovery transparent. We know it used algebraic number theory. We don’t fully understand why it reached for that tool.

What this does suggest is where AI’s most surprising leverage may come from: cross-domain synthesis. When a problem sits at the boundary between fields, human mathematicians — who specialise deeply and develop strong intuitions within one area — are less likely to look across to another. The model doesn’t carry those boundaries. That’s a different kind of reasoning resource, and this result is the clearest evidence yet of what it can produce.

Erdős reportedly offered $500 for a solution to this problem. The prize is posthumous. The mathematics is real.

References

  • OpenAI, “An OpenAI model has disproved a central conjecture in discrete geometry” (May 2026) — openai.com
  • Alon et al., “Remarks on the Disproof of the Unit Distance Conjecture” — arxiv.org/abs/2605.20695
  • OpenAI companion PDF — cdn.openai.com
  • Scientific American, “AI Just Solved an 80-Year-Old Erdős Problem and Mathematicians Are Amazed” — scientificamerican.com
  • Gil Kalai, “Amazing: Erdős’ Unit Distance Problem was Disproved!” — gilkalai.wordpress.com
  • Interesting Engineering, “80-year-old geometry puzzle cracked by OpenAI using number theory” — interestingengineering.com
  • Technology.org, “OpenAI Claims a Real Math First After Earlier Flop” — technology.org