OpenAI says its reasoning model has produced an original mathematical proof disproving a famous geometry conjecture that has stood unsolved since 1946. The problem, posed by legendary mathematician Paul Erdős, asked whether the best way to arrange points at unit distances resembles a square grid. For 80 years, mathematicians believed it did. OpenAI's model discovered an entirely new family of constructions that performs better — and this time, the mathematicians who exposed OpenAI's last embarrassing math claim are backing it up.
Why "For Real This Time" Matters
Seven months ago, OpenAI's former VP Kevin Weil posted on X that GPT-5 had solved 10 previously unsolved Erdős problems. It turned out the model had simply found existing solutions already in the literature. Rivals like Yann LeCun and Google DeepMind CEO Demis Hassabis mocked the claim. Thomas Bloom, who maintains the Erdős Problems website, called it a dramatic misrepresentation. Weil deleted his post. The incident became one of OpenAI's most embarrassing public moments.
This time, OpenAI published companion remarks from the same mathematicians who called out the previous failure. Bloom, alongside Noga Alon and Melanie Wood, confirmed the result is genuine. The proof came from a general-purpose reasoning model — not a system specifically designed to solve math problems.
The validation matters enormously. OpenAI's credibility on mathematical claims was damaged. Having the exact mathematicians who exposed the last failure endorse this result is the strongest possible confirmation.
What the Model Actually Did
The conjecture concerned unit-distance graphs — arrangements of points where every pair at distance exactly one is connected. Since 1946, the best known constructions resembled square grids. OpenAI's model discovered a new family of constructions that outperforms the grid approach — disproving the conjecture that grids were optimal.
OpenAI says this is the first time AI has autonomously solved a prominent open problem central to a field of mathematics. The model held together a long chain of reasoning, connected ideas across mathematical subfields, and explored approaches that human researchers had not considered.
The implications extend beyond geometry. If AI systems can produce original mathematical proofs — not by searching databases but by reasoning through novel approaches — the applications span drug discovery, physics, engineering, and materials science.
The Broader AI Research Race
The result arrives during a period of intense competition over AI reasoning capabilities. Google DeepMind has published its own mathematical AI results. Anthropic has focused Claude on enterprise applications rather than mathematical research. And Adaption's AutoScientist is automating AI training itself — a step toward AI systems that improve their own capabilities.
OpenAI's math breakthrough is specifically valuable because it demonstrates reasoning rather than pattern matching. AI chatbots that give wrong health advice do so because they generate statistically probable responses rather than reasoning through medical logic. A model that can construct an original mathematical proof is demonstrating a fundamentally different capability — one that could eventually address the reliability problems plaguing AI across every application.
What Mathematicians Think
Bloom's endorsement was characteristically measured. He said AI is helping to more fully explore the cathedral of mathematics built over centuries. He asked what other unseen wonders are waiting.
The framing is important. Bloom is not saying AI replaced a mathematician. He is saying it found something mathematicians missed. The model explored a possibility space that humans had not — not because humans could not, but because the search space was too large for manual exploration.
That distinction positions AI as a tool for mathematical discovery rather than a replacement for mathematicians. The model generates proofs. Humans verify them. The collaboration produces results that neither could achieve alone.
What It Means for OpenAI
The Erdős result gives OpenAI something it desperately needs: a genuine technical achievement at a moment when its enterprise market is being overtaken by Anthropic, its co-founders are leaving for competitors, and its trial just ended with a victory that was overshadowed by weeks of damaging testimony.
A verified mathematical breakthrough — endorsed by the researchers who called out OpenAI's last failure — is the kind of result that rebuilds credibility. It reminds the AI industry that OpenAI can still produce frontier research, even as its competitive position in enterprise and coding markets erodes.
Whether one mathematical proof is enough to reverse that trajectory is another question entirely.
