OpenAI Disproves 80-Year Erdős Math Conjecture
An internal OpenAI reasoning model produced an original proof disproving the Erdős unit distance conjecture, the first time AI autonomously solved a major open problem in mathematics.
5 min read
An internal OpenAI reasoning model has produced an original proof disproving the Erdős unit distance conjecture - a problem in combinatorial geometry that sat open since 1946. Fields Medalist Timothy Gowers said the result deserves publication in the Annals of Mathematics, one of the most selective journals in the discipline, "without any hesitation."
TL;DR
- OpenAI's general-purpose reasoning model disproved the Erdős unit distance conjecture through an original proof
- First time AI autonomously solved a prominent open problem central to a field of mathematics
- Timothy Gowers (Fields Medal, 1998) said the proof is publishable in Annals of Mathematics without hesitation
- The model wasn't trained for math - Noam Brown confirmed it's a general-purpose LLM
- 15 Erdős problems solved since January 2026; 11 credited to AI
The conjecture, first posed by Paul Erdős in 1946, asked how many point pairs in a n-point plane set can sit at unit distance from each other. For nearly 80 years, mathematicians assumed the best constructions looked like square grids. The OpenAI model found an entirely new family of constructions that performs better - a polynomial improvement over the grid baseline.
The Claims
| Claim | Source | Verdict |
|---|---|---|
| First AI to autonomously solve a major open math problem | OpenAI | Confirmed by external mathematicians |
| Square grids aren't ideal for unit distance problems | Model proof | Confirmed - new constructions yield polynomial improvement |
| Model is general-purpose, not math-specialized | Noam Brown | Confirmed publicly |
| Proof qualifies for Annals of Mathematics | Timothy Gowers | Confirmed - Fields Medalist stated "without any hesitation" |
"I would recommend this work for publication in the Annals of Mathematics without any hesitation."
- Timothy Gowers, Fields Medal winner (1998), Cambridge University
The Erdős unit distance problem dates to 1946 and asks how densely points can be packed in the plane while maximizing pairs at distance exactly one.
Source: pexels.com
What Happened
What They Measured
The planar unit distance problem is simple to state. Place n points anywhere in the plane. Count how many pairs sit at distance exactly 1 apart. The question: what placement maximizes that count?
Since Erdős's original work, the best lower bounds came from square lattice constructions. The proof OpenAI published constructs point sets by embedding norm-one elements into a high-dimensional Minkowski lattice, cutting by a product of discs, and projecting to one complex coordinate. The critical ingredients borrow from algebraic number theory - class field theory specifically - a field that normally has nothing to do with combinatorial geometry. That cross-disciplinary connection is what made the result hard to find by conventional search. Three external mathematicians reviewed and signed off on the proof: Noga Alon, Melanie Wood, and Thomas Bloom. Bloom, who maintains the Erdős Problems website and was vocal about OpenAI's false claims last year, wrote: "AI is helping us more fully explore the cathedral of mathematics we have built over the centuries."
The summarized chain of thought runs to 125 pages. The full underlying reasoning is longer.
What They Didn't
OpenAI hasn't named the specific model. Noam Brown stated it plainly: "This is a general-purpose LLM. It wasn't targeted at this problem or even at mathematics." That distinction is real. This isn't a math-specialized system like the IMO-targeted models Google DeepMind rolled out - it's the same class of model being used for code review and document summarization.
The proof also hasn't been verified inside a formal proof assistant. External mathematicians confirmed it, but machine-checkable verification in Lean 4 or Coq remains a separate step that hasn't been completed.
One more limit worth stating: the Erdős unit distance conjecture originally concerns the exact asymptotic behavior of the unit distance function. Disproving square grids as best is significant but doesn't close the conjecture completely. The result narrows the problem meaningfully, but the upper bound question remains open.
The proof connects algebraic number theory to discrete geometry - a cross-field application that conventional mathematical search hadn't produced in 80 years.
Source: pexels.com
A Year of Fast Moves
This result sits roughly 10 months after AI systems first hit gold-level performance at the 2025 International Mathematical Olympiad. In January 2026, research teams began moving through Erdős's open problem list at a pace that surprised most observers. Since then, 15 problems have moved from open to solved, with 11 of those credited specifically to AI models.
The qualitative difference matters. IMO problems are competition problems with existing solutions somewhere in the literature - exceptionally hard, but not open research. The Erdős unit distance problem was truly unresolved, with active professional work on it for eight decades. Terence Tao noted earlier this year that ideas were becoming cheap while verification remained expensive. This result moves one step past idea generation into producing a proof the community accepts as original.
Context for credibility: in October 2025, then-VP Kevin Weil claimed GPT-5 had solved 10 Erdős problems. It hadn't - the model had rediscovered existing solutions without producing anything new. OpenAI's credibility was damaged. This time, the company published the full proof document, named specific validators, and waited for a Fields Medalist's written endorsement before announcing. That's a different standard of evidence than the October 2025 rollout.
Should You Care?
If you run math research, yes - the same general-purpose model handling your literature searches can now produce publishable proofs in combinatorial geometry. That's a workflow change whether you want it.
For engineers, the concrete signal is what a 125-page cross-disciplinary chain of thought implies about the reasoning engine underneath. A model connecting algebraic number theory to discrete geometry without targeted training is the same category of system you use as a coding assistant. The gap between "useful tool" and "independent research contributor" is closing on a faster schedule than most roadmaps projected.
Noam Brown's comment is worth sitting with: "Less than 1 year ago frontier AI models were at IMO gold-level performance. I expect this pace of progress to continue."
In mathematics, disproving one class of optimal constructions is the step before resolving the asymptotic bound entirely. How long the next step takes is now a question the machine can help answer.
Sources:
- OpenAI: An OpenAI model has disproved a central conjecture in discrete geometry
- TechCrunch: OpenAI claims it solved an 80-year-old math problem - for real this time
- Noam Brown (@polynoamial) - Research Scientist, OpenAI
- Quanta Magazine: The AI Revolution in Math Has Arrived
- Digg: OpenAI says its internal general-purpose model solved the unit distance problem
