OpenAI Model Cracks 80-Year-Old Erdős Geometry Conjecture

OpenAI announced on Tuesday that one of its internal general-purpose reasoning models has independently disproved a central conjecture in discrete geometry that mathematicians have circled for nearly 80 years. The result, published with an accompanying proof paper and a set of remarks from independent mathematicians, marks the first time an AI system has autonomously settled a major open problem in a recognized field of mathematics.

The proof concerns the planar unit distance problem, a question Hungarian mathematician Paul Erdős first posed in 1946. The model's solution beat the previously assumed optimal arrangement and connected the geometry to algebraic number theory in a way that surprised the human experts who reviewed it.

What the conjecture asked

The unit distance problem sounds simple. Take n dots and place them on a flat surface. How many pairs of those dots can sit exactly one unit apart? Erdős asked whether the upper bound was tight against a specific function. For decades, the best known constructions were variations on the square grid, and most working mathematicians assumed grids were essentially the answer.

The model found a different family of constructions that beats the grid by a polynomial factor and proved the new bound rigorously. In plain terms, the AI did not just find one better example. It found an infinite family of better examples and explained why they work.

Why this matters: the problem sits at the intersection of combinatorics, geometry, and number theory. Progress here unlocks tools used in coding theory, computational geometry, and parts of theoretical computer science where unit-distance arguments show up as building blocks for harder results.

The mathematicians signed off

OpenAI did not publish in isolation. The company circulated the proof to three established researchers and published their independent remarks alongside the technical paper. Noga Alon of Princeton, one of the world's leading combinatorialists and the mathematician who described the unit-distance problem as one of Erdős's favorites, called the outcome "an outstanding achievement, settling a long-standing open problem."

Melanie Wood of Harvard and Thomas Bloom of Manchester each verified the proof and offered their own commentary. Fields Medalist Tim Gowers, separately, said he would recommend the proof for publication "without any hesitation" — a phrasing that matters because Gowers has spent the last several years writing skeptically about AI mathematical claims.

The independent verification is the difference between this announcement and OpenAI's previous claim in October 2025. That earlier claim, around a different problem, collapsed when outside experts found gaps in the reasoning. This time the company waited until the proof had been audited by people who could read it.

How the model got there

The technical surprise is not just the result. It's the method. The model attacked a geometry problem using algebraic number theory, an entirely different branch of math. Specifically, it used properties of certain number fields to construct point configurations that satisfy the geometric constraint more efficiently than any grid.

That kind of cross-domain connection is the sort of move humans get credit for when they pull it off. It's what mathematicians mean when they talk about "deep" proofs. The model produced the connection without being prompted toward it, and the resulting argument is short enough to read in one sitting.

OpenAI emphasized that this was not a math-specialized model. It's a general-purpose reasoning system, the same lineage that handles code, planning, and ordinary chat. The company has not disclosed which model version produced the proof or how many attempts it required.

Industry implications

For AI research, the result is a credibility milestone. Many of the field's biggest claims over the last two years have been about reasoning ability — that models can hold together long chains of logic, connect distant ideas, and arrive at non-obvious conclusions. The Erdős proof is a clean public example of all three, audited by skeptics.

For mathematics, the implications are messier. Working mathematicians will now have to decide whether to treat AI-assisted proofs as a tool, a co-author, or a threat to the discipline's identity. Journals will need to adapt review standards. Funding bodies will face questions about which problems are still worth assigning to humans.

For the rest of the AI industry, the lesson is competitive. OpenAI delivered a result with a verifiable, public artifact: a proof that either holds or doesn't. Most AI demos can be hand-waved. This one cannot.

What insiders are saying

Gowers's endorsement is the headline reaction, but the most telling response is Alon's framing. By calling it an Erdős favorite, Alon places the problem inside the canon of questions that working mathematicians cared about for personal reasons, not just technical ones. That signals the result is being taken seriously inside the field, not just inside AI circles.

Several mathematicians on X noted that the proof's brevity is itself notable. AI-generated proofs in the past have tended to be long, computer-checked artifacts that humans struggle to read. This one is short, human-legible, and uses techniques mathematicians already trust.

What's next

OpenAI has not announced which model produced the proof or whether it will be released. The company also has not committed to working through other open problems publicly. The proof itself and the accompanying remarks are available as PDFs on OpenAI's site, and the company has invited the math community to review the work in detail.

Expect more attempts soon. Anthropic, Google DeepMind, and a handful of well-funded labs all have reasoning programs targeting math benchmarks, and the public credibility from a verified open-problem proof is the kind of milestone that gets internally prioritized fast.

The bottom line

For 80 years, a famous question sat open. An AI answered it, mathematicians checked the work, and the answer holds. Whether or not the model can do this again on demand, the precedent is set. AI is now part of the publication record in pure mathematics, and that genie is not going back in the bottle.