OpenAI Reasoning Model Disproves 80-Year-Old Erdős Conjecture

On May 20, 2026, the global scientific community witnessed a watershed moment in both mathematics and computer science. OpenAI announced that an unreleased, internal general-purpose reasoning model had autonomously disproved the planar unit distance conjecture, an eighty-year-old open problem first posed by legendary Hungarian mathematician Paul Erdős in 1946. By constructing an elegant, highly original 125-page proof that successfully refuted the long-held Erdős conjecture, this artificial intelligence accomplished what decades of human mathematical brilliance could not. Rather than acting as a mere digital assistant, the system operated as an autonomous pioneer, fundamentally altering our understanding of discrete geometry and delivering a decisive blow to the skepticism that has long surrounded the reasoning limitations of large language models (LLMs). This achievement marks the first time a non-domain-specific neural network has solved a premier, long-standing open problem in pure mathematics.

The Geometry of Points and the Erdős Conjecture

To appreciate the scale of this breakthrough, one must understand the deceptively simple nature of the planar unit distance problem. Posed by Paul Erdős in his seminal 1946 paper, the problem asks a straightforward question: if you place n points on a flat, two-dimensional plane, what is the maximum number of pairs of points that can be exactly one unit of distance apart?

For generations, the prevailing mathematical consensus was governed by geometric intuition. Erdős himself observed that a simple square grid of points—properly scaled—could produce a high density of unit distances. Based on these grid configurations, he formulated the Erdős conjecture, proposing that the maximum number of unit-distance pairs, denoted as ν(n), could grow no faster than:

ν(n) ≤ n^{1 + c / log log n}

This upper bound, which translates asymptotically to n^1+o(1), represents a growth rate that rises only marginally faster than a purely linear rate. For eight decades, this conjecture stood as a foundational pillar of discrete geometry. While mathematicians succeeded in establishing lower and upper bounds—most notably the O(n^4/3) upper bound proven by Joel Spencer, Endre Szemerédi, and William Trotter in 1984—no one could find a point configuration that defied Erdős’s nearly-linear limit. The standard mathematical toolkit, heavily reliant on rescaled grids and classical geometric configurations, had reached an impasse, leaving researchers to believe that the limits of the physical plane simply forbade denser arrangements.

Bridging Geometry and Algebraic Number Theory

The OpenAI reasoning model shattered this eighty-year-old paradigm by completely abandoning the geometric avenues that had occupied human researchers for decades. Instead of attempting to optimize physical grid layouts, the AI recognized a profound, hidden connection between discrete planar geometry and deep algebraic number theory.

At the heart of the model’s proof is a masterclass in modern algebra, utilizing the Golod-Shafarevich criterion—a landmark 1964 theorem in non-abelian class field theory—to construct an infinite family of planar point configurations. The mathematical pipeline devised by the model can be broken down into several highly complex steps:

• Infinite Galois Towers: The model constructed an infinite, unramified tower of totally real number fields with 3-power Galois groups of growing degree. Within this tower, a carefully selected, fixed set of rational primes was forced to split completely.
• Adjoining Complex Units: By adjoining the imaginary unit i to these fields, the model generated high-dimensional lattices. These lattices contained an extraordinary abundance of elements whose images under every complex embedding possessed an absolute value of exactly one.
• Controlling Growth: Using Golod-Shafarevich theory, the model proved the existence of an infinite tower even after executing a quotient step to render the prescribed Frobenius classes trivial. This ensured that the resulting discriminants and class numbers grew at most exponential relative to the extension degree.
• Planar Projection: Finally, by projecting these multidimensional lattice points onto the complex plane, the model demonstrated the existence of point sets where the number of unit-distance pairs grew at a polynomial rate of n^1+ε.

Following the release of the proof, Princeton mathematician Will Sawin meticulously analyzed the model’s construction. He calculated the exact polynomial gain, determining the exponent improvement to be ε ≈ 0.014. By proving that the number of unit distances can scale at a rate of roughly n^1.014, the model successfully refuted the Erdős conjecture and demonstrated that algebraic constructions in number fields can systematically outperform traditional geometric grids.

Redemption and Validation by the Harshest Critics

In the world of pure mathematics, claims of solving legendary problems are met with extreme scrutiny. This skepticism was doubly intense for OpenAI. In October 2025, Kevin Weil, then leading the OpenAI for Science division, hastily announced on social media that a frontier model had solved ten unsolved Erdős puzzles. However, Thomas Bloom, a University of Manchester mathematician who maintains the authoritative erdosproblems.com registry, quickly exposed the claim as a “dramatic misrepresentation”. The model had not generated new mathematics; it had merely retrieved existing solutions from historical literature and framed them as novel discoveries.

The May 2026 announcement, however, was fundamentally different. Rather than relying on corporate marketing, OpenAI pre-empted criticism by releasing the raw 125-page proof alongside a comprehensive companion verification paper. Most remarkably, the co-authors of this verification paper included nine of the world’s most prominent mathematicians—including Thomas Bloom himself.

The verification team also featured Fields Medalist Timothy Gowers, who praised the elegance and logical robustness of the AI’s proof. Gowers went so far as to state that he would recommend the paper for publication in the prestigious Annals of Mathematics without hesitation. The co-authorship of elite minds such as Noga Alon, Daniel Litt, Arul Shankar, Jacob Tsimerman, Victor Wang, and Melanie Matchett Wood provided an absolute, ironclad seal of scientific legitimacy, transforming what could have been another public relations blunder into a historic milestone.

The Death of the “Stochastic Parrot” Narrative

For years, a vocal contingent of computer scientists and linguists argued that large language models were merely “stochastic parrots”—glorified statistical engines capable of compressing and recombining human text, but fundamentally incapable of genuine reasoning, creative problem-solving, or expanding the boundaries of human knowledge. Critics asserted that because neural networks are trained on historical data, they could never transcend their training distributions to produce truly novel, complex ideas.

The resolution of this discrete geometry problem forcefully dismantles that paradigm. The OpenAI model did not merely interpolate between existing papers; it synthesized a highly sophisticated bridge between two disparate fields of mathematics—algebraic number theory and discrete geometry—that human experts had never thought to connect in this specific manner.

Furthermore, the proof was generated in a “one-shot” pipeline, meaning the model produced the core mathematical architecture autonomously, without iterative human prompting, scaffolding, or step-by-step guidance. Leading number theorist Arul Shankar noted that this achievement marks a profound transition in the field of artificial intelligence: LLMs are officially shifting from passive research assistants that format code or search literature into active, autonomous intellectual entities capable of executing world-class scientific breakthroughs.

A New Frontier for AI and Human Mathematics

As the dust settles on this historic discovery, the broader implications for the future of scientific inquiry are staggering. This milestone represents a practical demonstration of how general-purpose reasoning models can serve as engines of discovery. By utilizing advanced chain-of-thought processing to break down monumental problems into granular, logically sound sub-steps, AI systems are now capable of navigating complex intellectual mazes that have stymied human minds for generations.

While the ultimate resolution of the planar unit distance problem remains open—since the AI did not establish a new absolute upper bound, but rather proved that the 80-year-old lower limit was incorrect—the methodology it introduced has completely redrawn the mathematical map. Mathematicians are now eagerly exploring the “number-field loophole” exposed by the model to see if other classical problems in discrete geometry, such as the Hadwiger-Nelson problem, can be unlocked using similar algebraic structures.

Ultimately, the disproof of the Erdős conjecture is more than just a victory for mathematics; it is a preview of a future where human intellect and machine reasoning merge to conquer the unsolvable. We have entered an era where the next great scientific revolution may not be authored by a human holding a chalk, but by a machine quietly processing a prompt.