On May 20, 2026, OpenAI announced that its new reasoning model autonomously disproved the Erdős unit-distance conjecture — an open problem in discrete geometry since 1946. Unlike the GPT-5 embarrassment seven months earlier, this time the proof has been independently verified by mathematicians including Thomas Bloom and Noga Alon.
Key takeaways
- OpenAI's model disproved Erdős's 1946 conjecture about unit-distance graphs
- Proof confirmed by Noga Alon, Melanie Wood, and Thomas Bloom (curator of erdosproblems.com)
- First time AI has autonomously solved a prominent open problem central to a field of mathematics — per OpenAI
- In October 2025 OpenAI falsely claimed GPT-5 solved 10 Erdős problems — those were existing solutions found in literature
- The proof came from a general-purpose reasoning model, not a math-specific system
What the Erdős conjecture is and why it stayed open for 80 years
In plain English: imagine you scatter 100 coins on a table. How many pairs of coins can you find that are exactly 1 cm apart? In 1946 Erdős asked: for n points in the plane, what is the maximum possible number of such pairs — and how do you arrange the points to achieve it?
For 80 years mathematicians believed the best layouts looked like a square grid — like graph-paper. OpenAI's model found a completely different way of arranging the points that produces even more unit-distance pairs, overturning the long-standing intuition.
Paul Erdős posed his unit-distance conjecture in 1946, asking about the maximum number of unit-distance pairs in a set of n points in the plane. For nearly 80 years, mathematicians assumed optimal configurations resembled square grids. OpenAI says its model discovered an entirely new family of constructions that outperform those grids — disproving the dominant intuition about what optimal solutions look like.
The Erdős conjecture belongs to a class of combinatorial problems requiring long chains of reasoning and connections across areas of mathematics. Solving it demands the ability to generate and formally verify extended proofs, not just search a numerical solution space.
After the GPT-5 lesson: this time verification came first
In October 2025, then-OpenAI VP Kevin Weil posted on X that GPT-5 had solved 10 Erdős problems. It turned out the model had merely found existing solutions in the literature. Taunts from rivals including Yann LeCun and Demis Hassabis followed quickly, and Weil deleted the post. Thomas Bloom, who runs erdosproblems.com and called that announcement "a dramatic misrepresentation," now explicitly endorses the current result.
"AI is helping us to more fully explore the cathedral of mathematics we have built over the centuries. What other unseen wonders are waiting in the wings?"
— Thomas Bloom, statement included with the OpenAI announcement
Noga Alon, one of the world's leading combinatorialists, and Melanie Wood, a MacArthur Fellowship recipient, have also confirmed the validity of the proof. OpenAI published a companion PDF with remarks from the mathematicians.
Architecture: a general model, not a specialized calculator
OpenAI emphasizes this is not a system built specifically for mathematics. It is a new general-purpose reasoning model — an architecture extended with long chain-of-thought capabilities that OpenAI has been developing for over a year.
The model had to explore a conceptual space, propose a new family of constructions, and formally verify them without relying on pre-existing solutions. The ability to connect ideas across fields — as OpenAI describes it — is precisely what distinguishes this result from the October mistake.
Broader context: AI as a tool for scientific discovery
Pure mathematics is one of the hardest tests for AI reasoning capability — a proof is either correct or it is not. In 2024, Google DeepMind demonstrated AlphaProof on International Mathematical Olympiad problems — but those were structured competition problems with known solution techniques.
OpenAI's result is different: the Erdős conjecture had no prior template solution to locate. The model had to generate something genuinely new. That places this result closer to scientific discovery than efficient optimization. The key question for the mathematical community is whether the published proof will pass full peer review in a refereed journal.
Why it matters
The sharpest long-standing criticism of language models in mathematics was this: LLMs cannot generate original proofs — they can only recognize and rephrase existing ones. The May 20, 2026 result, if it survives full review, is the first clear answer to that criticism.
The implications extend beyond discrete geometry. If a general reasoning model — without pre-loaded domain expertise — can overturn an 80-year mathematical assumption, it opens questions about applying similar systems in computational biology, theoretical physics, and cryptography. OpenAI explicitly names these fields as potential next targets.
One important caveat: OpenAI's announcement describes a discovery, not a product. The model is not publicly available as a mathematical tool. The pace of commercialization and integration into research workflows is a separate, still-open question.
What's next
- OpenAI announced the full formal proof is ready for independent verification by the mathematical community
- Thomas Bloom stated the erdosproblems.com site will update the problem's status after formal review concludes
- OpenAI identified biology, physics, and engineering as next areas to test reasoning AI as a discovery tool
Sources
- OpenAI Blog — OpenAI model disproves discrete geometry conjecture
- TechCrunch — OpenAI claims it solved an 80-year-old math problem — for real this time
- The Erdős Problems — erdosproblems.com
- OpenAI — Unit Distance Remarks (PDF)