OpenAI Disproved an Erdős Conjecture, and the Math Checks Out

Seven months ago, OpenAI's math claims were a punchline. An executive announced that GPT-5 had "solved" ten open Erdős problems; it turned out the model had found existing solutions in the literature for problems mislabeled as open, the post got deleted, and Thomas Bloom — who maintains the Erdős problems database the claim was built on — called it a dramatic misrepresentation.

Now OpenAI says an internal reasoning model has disproved the Erdős unit distance conjecture, a problem open since 1946. The reflex is skepticism. Resist it this time: the result shipped with an explicit bound, a follow-up paper by Princeton's Will Sawin, and a companion write-up co-authored by nine mathematicians including Noga Alon, Timothy Gowers, and Melanie Matchett Wood. Gowers — a Fields Medalist, not a man given to AI boosterism — told reporters he'd recommend the proof to the Annals of Mathematics without hesitation. The math checks out, and that changes what kind of story this is.

What the model actually found

Erdős asked a deceptively simple question: place n points in the plane — how many pairs can sit at exactly distance 1? Cleverly scaled grids give roughly n^(1 + c/log log n) unit distances, barely more than linear, and for 80 years the working assumption was that grids were essentially optimal. The best known upper bound, n^(4/3), has stood since 1984, and most effort went into closing the gap from above.

The model attacked from below. It built point configurations from rings of algebraic integers — lattices in higher dimensions, projected down to the plane — pulling in machinery from deep algebraic number theory, reportedly including infinite class field towers. That's territory almost nobody working on the unit distance problem touches. Sawin then made the improvement explicit: the construction achieves at least n^1.014 unit distances, a genuine polynomial beat over grids. Conjecture dead.

Keep the asterisks in view. The improvement only kicks in at around 10^2,000,000 points — this is an existence result, not usable geometry. It's a disproof by construction, which is search-shaped work, not theory-building. And the gap to n^(4/3) remains wide open. None of that weakens the result's validity; counterexamples are how conjectures die. But it does tell you which capability was actually demonstrated.

Breadth plus grind, not genius

Gowers noted that many of the ingredients were already in the literature. Jacob Tsimerman, one of the nine verifying authors, said he'd considered a similar approach himself and abandoned it — that kind of attack eats months and usually fails, so no rational human career invests in it.

That's the honest shape of this breakthrough, and it should look familiar to anyone who uses these models for code. The model didn't invent new mathematics. It combined two fields that almost no single human holds simultaneously at research depth, and it ground through a low-probability strategy without the disincentives that make humans quit. Per Understanding AI's reporting, even the internal model only succeeded about half the time at maximum reasoning budget. This is the "obscure library you'd never have found" experience, scaled up to the research frontier — and it turns out the research frontier has more retrieval-and-synthesis-shaped problems than mathematicians assumed.

That framing also punctures the rumor mill. Claims circulated that an OpenAI model family called "Astra" had knocked out ten major open problems. OpenAI has confirmed nothing of the sort — there is exactly one validated result. Meanwhile Google DeepMind announced solutions to nine Erdős problems two days after OpenAI's post, which tells you the labs have discovered that Bloom's database is the rare benchmark you can't juice: the answers are checkable and the problems are certifiably open. Expect it to become a leaderboard. That's healthier than most eval races.

The bottleneck moved, and that's the developer story

What actually separated May 2026 from October 2025 wasn't primarily model capability — it was process. The result went out with named verifiers, a cleaned-up human-checked proof, and an arXiv trail. A team of elite mathematicians spent real effort validating, tightening, and extending raw model output before anyone claimed anything.

Generalize that and you get the near-term playbook for AI-assisted research of any kind: generation is now cheap and unreliable, so verification is the scarce resource. In math, that pushes toward formal tools — a Cambridge undergraduate reportedly got the first autonomous Erdős problem solution in January by pairing a public model with Harmonic's Aristotle prover, where the proof checker plays the role the nine mathematicians played here, at zero marginal cost. In software, it's the same argument you already know: models that write code are only as useful as the test harness that catches them lying. The labs just demonstrated that principle at the level of the Annals of Mathematics.

The other detail worth sitting with: Understanding AI reports that GPT-5.5 — a model you can call today — can reproduce the disproof given light hints. If that's right, publicly available frontier models plausibly contain results nobody has harvested yet, because nobody pointed them at the right question with enough compute and a verification pipeline on the other end. The binding constraints are problem selection and checking, not model access. For anyone working in fields with mechanically checkable answers — combinatorics, algorithms, parameter optimization — that's an invitation, though the cost profile is brutal: a coin-flip success rate at maximum token budget is a research-lab workflow, not an API integration.

Real, narrow, and a preview

Verdict: this is a genuine shift, not hype — the first autonomous AI result that mathematicians say is exciting on its own merits rather than as a leading indicator, as Daniel Litt put it. It's also narrower than the headlines: one counterexample, built from existing ideas, by an unreleased model, at heroic compute cost, with humans still doing all the trust-establishment. Gowers's own read is the one to keep — he expects models to reach into theory-building and question-asking next. Until then, the tell that separates real claims from the next Astra rumor is simple: look for the verifiers' names and the arXiv link. This one had both.