OpenAI Reasoning Model Solves 80-Year-Old Erdős Math Problem

An internal artificial intelligence model developed by OpenAI has autonomously solved an 80-year-old mathematical mystery, disproving a long-held conjecture by the legendary Hungarian mathematician Paul Erdős. The discovery, announced in late May 2026, has sent ripples through the mathematical and technological communities, marking one of the first times a machine has generated a genuinely original proof for a prominent open problem.[1][2]

The mystery in question is known as the "planar unit distance problem," first posed by Erdős in 1946. The premise is deceptively simple: if you place a certain number of dots on a flat sheet of paper, what is the maximum number of pairs that can be exactly one unit of distance apart? Erdős conjectured that the number of pairs would rise only slightly faster than the number of dots themselves.[2][3]

For nearly eight decades, the mathematical consensus was that the optimal arrangement to maximize these unit distances would look roughly like a square grid. Generations of mathematicians attempted to prove or disprove this intuition, but the sheer complexity of calculating the exact upper bounds for large numbers of points left the problem unresolved.[1][5]

That changed on May 20, 2026, when OpenAI revealed that its general-purpose reasoning model had shattered this assumption. The AI discovered an entirely new, infinite family of point arrangements that outperformed the traditional grid-based constructions, definitively proving that Erdős's proposed limit was too low.[1][4]

What makes the achievement remarkable is not just the result, but the method. Rather than brute-forcing millions of geometric combinations, the AI achieved the breakthrough by connecting discrete geometry with deep algebraic number theory. This cross-disciplinary leap—borrowing complex tools from one distant field of mathematics to solve a problem in another—had eluded human experts for 80 years.[4][6]

The model used was not a specialized calculator trained exclusively on mathematical formulas. According to OpenAI researcher Sébastien Bubeck, it was a general-purpose large language model trained for advanced reasoning. The researchers simply provided the model with a prompt describing the conjecture and instructed it to either prove or disprove it, without guiding it down any specific mathematical path.[3][4]

The mathematical community quickly moved to validate the machine's work. Fields Medalist Tim Gowers hailed the result as a "milestone in AI mathematics," while Princeton combinatorialist Noga Alon described the solution as an "outstanding achievement" that applied sophisticated algebraic tools in an elegant and clever way.[4][5]

Despite the AI's autonomous heavy lifting, human mathematicians still played a crucial role in the final output. Thomas Bloom, a mathematician who maintains a database of Erdős problems and had previously criticized OpenAI for a premature claim, co-authored a companion paper validating this new breakthrough. Bloom noted that while the AI's original proof was completely valid, human researchers significantly improved, digested, and extended the proof to explore its broader consequences.[2][3]

It is important to note that the AI did not completely solve the entirety of the unit distance problem. While it successfully disproved the square-grid conjecture and showed that Erdős's limit was too low, the exact rate at which the pairs of dots rise—the true upper bound—remains an open question for future research.[2][5]

Nevertheless, the philosophical implications of the breakthrough are profound. For years, critics have dismissed large language models as "stochastic parrots" that merely retrieve and remix existing text from their training data. Because this specific proof did not exist anywhere in mathematical literature, the model had to genuinely synthesize new ideas and reason its way to a novel conclusion.[5][6]

OpenAI's success rate also highlights the iterative nature of AI reasoning. The research team ran their prompt through the model multiple times, and it produced the correct, verified solution in roughly 50 percent of those trials, demonstrating that the model was actively exploring different logical pathways rather than reciting a memorized answer.[3]

The success has prompted a broader conversation about the future of scientific discovery. In early June, a group of experts published the Leiden Declaration, calling for new guardrails and standards around the use of AI in mathematical research to ensure proofs can be properly audited.[3]

If an artificial intelligence can autonomously connect distant fields of mathematics to solve an 80-year-old geometry problem, researchers believe the same underlying reasoning capabilities could soon be deployed elsewhere. The tech industry is now looking toward a near future where AI acts as a collaborative partner in accelerating breakthroughs in materials science, drug development, and climate engineering.[1][6]