May 21, 2026 — In a major milestone for artificial intelligence and mathematics, OpenAI announced that one of its internal general-purpose reasoning models has autonomously disproved a central conjecture in discrete geometry first posed by legendary mathematician Paul Erdős in 1946.
This marks the first time an AI system has independently solved a prominent, long-standing open problem at the heart of an active mathematical field, according to OpenAI.
The Problem: How Many Unit Distances in the Plane?
The planar unit distance problem asks: Given n points in the Euclidean plane, what is the maximum number of pairs that can be exactly distance 1 apart?
- Erdős offered a monetary prize for progress and called it one of his favorites.
- Simple constructions like points on a line give ~n pairs.
- The classic square lattice (grid) gives roughly n × (some slow-growing factor).
- For decades, mathematicians believed the best possible was on the order of n^{1 + o(1)} — slightly more than linear, with the extra term vanishing as n grows. This was widely viewed as essentially optimal.
The best known upper bound remains O(n^{4/3}) from 1984 (Spencer–Szemerédi–Trotter), but the conjecture focused on the lower-bound constructions.
The Breakthrough: A New Family of Constructions
OpenAI’s model discovered an infinite family of point configurations that achieve n^{1 + δ} unit distances for some fixed δ > 0 (a polynomial improvement). A later refinement by mathematician Will Sawin shows δ can be at least 0.014.
Key innovation: The proof draws on sophisticated tools from algebraic number theory (class field towers, Golod–Shafarevich theory, and generalizations of Gaussian integers) applied to an elementary geometric question. This unexpected bridge surprised experts.
The model produced a full proof (originally ~125 pages of chain-of-thought), which external mathematicians verified, simplified, and expanded in a companion paper. The proof is considered strong enough for submission to top journals like the Annals of Mathematics.
Proof and companion materials (from OpenAI):
- Main proof PDF
- Companion remarks by mathematicians
- Abridged chain-of-thought
Expert Reactions: “A Milestone in AI Mathematics”
Leading mathematicians have praised the result:
- Noga Alon (Princeton): Called it an “outstanding achievement” and noted the surprise that the answer exceeds the long-believed bound.
- Fields Medalist Tim Gowers: Described it as “a milestone in AI mathematics” and said he would recommend acceptance without hesitation.
- Arul Shankar: “Current AI models go beyond just helpers… they are capable of having original ingenious ideas.”
- Thomas Bloom (maintainer of erdosproblems.com): Highlighted that it teaches us something new — number-theoretic constructions have more power than suspected.
This comes after a previous OpenAI math claim in late 2025 was criticized for overstating “solutions” that were actually rediscoveries of known results. Experts note this time the proof is original and rigorously verified.
Comparison with Other AI Math Models & Achievements
| Achievement | Model / System | Year | Type | Significance |
|---|---|---|---|---|
| Unit Distance Disproof | OpenAI internal reasoning model | 2026 | General-purpose, autonomous | First AI solving prominent open research problem in a subfield; novel cross-field ideas |
| IMO Gold Medal Level | OpenAI & Google models | 2025 | Reasoning models | Solved contest problems at human gold level; not open research |
| AlphaGeometry / AlphaProof | Google DeepMind | 2024–2025 | Specialized geometry/proof systems | IMO silver/gold on formalized problems; domain-specific |
| Erdős Primitive Set (amateur) | GPT-5.4 Pro via user | 2026 | Prompted general model | Solved 60-year problem with new method |
| FrontierMath Benchmark | OpenAI o3 series | 2025–2026 | Reasoning models | Improving but still limited on hardest problems |
What makes this unique:
- It used a general-purpose model (not math-specialized or scaffolded).
- It made an original discovery with unexpected techniques from another field.
- The result advances human knowledge rather than just matching existing proofs.
Previous systems like AlphaGeometry excelled at formalized Olympiad problems but hadn’t cracked open research conjectures autonomously.
Why This Matters Beyond Math
Stronger long-horizon reasoning and cross-domain connections could accelerate progress in physics, biology, materials science, and AI research itself. OpenAI frames this as early evidence of AI becoming a true collaborative partner in frontier science — one that can explore “spiky” frontiers humans might overlook.
Human expertise remains essential for choosing problems, interpreting results, and refining ideas. But the “cathedral of mathematics” just got a powerful new explorer.
What’s next? Mathematicians are already exploring whether similar number-theoretic techniques apply to other open problems in discrete geometry and beyond.
This breakthrough reinforces 2026 as a pivotal year in AI’s reasoning capabilities — moving from solving exercises to contributing genuine discoveries.
Leave a Comment