OpenAI’s AI autonomously disproves 80-year-old Erdős conjecture, marking first frontier mathematical discovery
A reasoning model independently solved the Erdős unit distance problem in May 2026, validating AI as a genuine research tool in pure mathematics and accelerating the timeline for machine-driven scientific discovery.
OpenAI announced on 20 May 2026 that an internal reasoning model autonomously disproved the Erdős unit distance conjecture, an 80-year-old problem in discrete geometry first posed in 1946.
The proof spans 125 pages across two companion papers, according to hitechies.com, and has been validated by leading mathematicians including Fields Medallist Tim Gowers. This marks the first time artificial intelligence has independently solved a prominent research-frontier problem central to a mathematical field, not merely retrieved an existing solution from literature.
The breakthrough arrives after OpenAI’s credibility took a hit in October 2025, when the company claimed its GPT-5 model had solved ten Erdős problems—a claim later revealed to have simply found existing solutions already in published work. Mathematician Thomas Bloom called that episode a “dramatic misrepresentation”, per roborhythms.com. The May 2026 result underwent independent verification by top-tier mathematicians including Noga Alon, Will Sawin, and Arul Shankar to establish its legitimacy.
The mathematical significance
The Erdős unit distance conjecture asks: what is the maximum number of unit-distance pairs among n points in a plane? For eight decades, mathematicians believed square grids were essentially optimal, achieving approximately n^(4/3) unit-distance pairs. The AI discovered constructions using algebraic number theory—specifically the Golod-Shafarevich criterion and infinite class field towers—that achieve polynomial improvement over square grids, explainx.ai reports. Princeton mathematician Will Sawin subsequently refined the result to n^(1.014) unit-distance pairs.
“It feels like magic. It’s kind of an amazing experience to have a machine give back something which really resembles how I work.”
— Mathematician quoted in Scientific American
The model produced hundreds of pages of calculations using chain-of-thought reasoning rather than brute-force verification, systematically exploring paths that human mathematicians had dismissed, according to Scientific American. Canadian mathematician Daniel Litt described it as “the first result produced autonomously by an AI that I find interesting in itself.”
Autonomy and architecture
OpenAI’s claim of autonomy—that the model required minimal human intervention beyond the initial problem prompt—distinguishes this achievement from prior computer-assisted proofs. The system connected discrete geometry to algebraic number theory, a cross-domain leap that Singularity Hub notes required virtually no human guidance during the reasoning process itself.
Paul Erdős posed over 1,500 problems during his lifetime, many with monetary prizes attached. The unit distance conjecture represented a central question in combinatorial geometry, with implications for distance graph theory and computational complexity. Previous attempts focused on refining grid-based constructions rather than exploring algebraic structures.
Fields Medallist Tim Gowers called the work “a milestone in AI Mathematics”, per pasqualepillitteri.it. The validation process involved multiple independent mathematicians verifying both the technical correctness and the novelty of the approach. As of late May 2026, the proof has been reviewed by specialists but awaits formal peer-reviewed journal publication, according to Nature.
Competitive acceleration
Two days after OpenAI’s announcement, Google revealed on 22 May that its AI system had solved nine open Erdős problems, including two open for over 50 years, understandingai.org reports. The timing suggests intensifying competition among frontier labs to demonstrate reasoning capabilities beyond language processing.
The breakthrough validates reasoning-based architectures that decompose complex problems into verifiable steps, a departure from the pattern-matching approaches that dominated earlier AI mathematics efforts. OpenAI has not disclosed which internal model produced the result or whether it will be released publicly.
Implications for mathematical research
The result forces a reckoning about the future role of human mathematicians in specialised domains. Unlike theorem-proving systems that verify human-designed proofs, this model generated novel constructions without a predetermined path. The algebraic number theory approach—connecting Golod-Shafarevich towers to geometric point configurations—represents the kind of conceptual leap that defines creative mathematical work.
- First autonomous AI solution to a prominent open problem in pure mathematics, not retrieval of existing work
- Proof validated by Fields Medallists and specialists, awaiting peer-reviewed publication
- Chain-of-thought reasoning produced cross-domain connections human mathematicians had missed
- Google announced competing results two days later, signaling accelerated timeline for AI-driven discovery
Experts quoted in OpenAI materials emphasise that “expertise becomes more valuable, not less” in an AI-assisted research environment, with humans choosing problems, interpreting results, and directing inquiry. The distinction matters: this model did not select the Erdős conjecture autonomously but required human researchers to frame the problem and evaluate whether the output constituted a meaningful solution.
The credibility stakes for OpenAI remain high after the October 2025 misstep. TechCrunch notes that independent verification by named mathematicians Noga Alon, David Wood, and Thomas Bloom was essential to establishing legitimacy this time.
What to watch
Full peer-reviewed publication of the proof will provide technical scrutiny that external mathematician validation cannot fully replace. The competitive response from Google and other frontier labs suggests mathematical reasoning will become a key benchmark for evaluating AI capabilities beyond language tasks. Whether OpenAI releases the model publicly—or keeps it internal as a research tool—will signal how the company views the commercial versus scientific value of autonomous mathematical reasoning. The timeline for AI assistance in adjacent fields like theoretical physics and algorithm design has likely compressed, with the first autonomous contributions arriving sooner than most researchers expected a year ago.