OpenAI says a general-purpose reasoning model found a new construction for the planar unit distance problem, disproving a belief dating back to Erdős's 1946 question.
Last checked: May 21, 2026. This article is based on OpenAI's official announcement, the proof released by OpenAI, the companion remarks paper and the abridged chain-of-thought material linked from the company post.
OpenAI says one of its general-purpose reasoning models has disproved a longstanding conjecture in discrete geometry, finding a new family of point configurations for the planar unit distance problem that beats the square-grid-style constructions mathematicians had believed were essentially best possible for nearly 80 years.
The company announced the result on May 20 in a research post titled "A model disproves a central conjecture in discrete geometry." OpenAI framed it as a milestone for both mathematics and AI, saying the proof came from a general-purpose reasoning model rather than a system built specifically for this problem or for mathematics competitions.
At the center of the result is the planar unit distance problem, a question first posed by Paul Erdős in 1946. The problem asks how many pairs of points at distance exactly 1 can be formed among n points in the Euclidean plane. For decades, the best known constructions looked roughly like square grids, and many mathematicians believed that this was close to the right answer. OpenAI says its model found a construction that performs better by a polynomial factor for infinitely many values of n.
The claim is important for two reasons. Mathematically, it changes the landscape around a central problem in discrete geometry. Technically, it gives a concrete example of an AI system producing a new proof on a prominent open problem rather than only assisting with search, exposition or proof checking.
OpenAI said the proof was checked by a group of external mathematicians and released three official materials: the proof itself, a companion paper with remarks and context from mathematicians, and an abridged chain of thought showing part of how the model approached the problem.
What OpenAI announced
OpenAI's announcement says the model discovered an entirely new family of constructions for the planar unit distance problem. The result disproves the belief that the best possible solutions must look essentially like square grids.
The company describes the result as the first time AI has autonomously solved a prominent open problem central to a field of mathematics. That wording is significant. AI systems have previously helped mathematicians with search, conjecture generation, formalization and proof assistance. OpenAI is claiming something narrower and stronger here: the core mathematical proof came from a general-purpose reasoning model working on a real open problem.
The model was not presented as a specialized geometry solver. According to OpenAI, it was a general-purpose reasoning model trained for broad problem solving. The company said it evaluated the model on a collection of Erdős-style open problems and that the system produced a proof for the planar unit distance problem without being specifically trained or scaffolded for that problem.
OpenAI also released supporting documentation. The proof PDF presents the construction and argument. The companion remarks paper, written with several mathematicians, places the result in context and explains how the construction connects to earlier work. The abridged chain-of-thought document shows selected model reasoning in a shortened form.
That collection of materials matters because the announcement is not only a product claim. It is also a mathematical claim that requires inspection by experts. OpenAI said external mathematicians checked the proof, and the companion document records reactions and explanatory notes from researchers who reviewed the result.
OpenAI also said it studied how often its models succeeded on this problem under different levels of test-time compute after the initial proof was verified. The company did not frame the breakthrough as a one-off benchmark score; it presented the result as part of a broader evaluation of whether advanced models can contribute to frontier research.
The planar unit distance problem
The planar unit distance problem is easy to state and hard to solve. Place n points in the plane. Count how many pairs of those points are exactly distance 1 apart. The question is: how large can that number be?
This maximum is usually written as u(n). A line of n equally spaced points gives n - 1 unit distances. A square grid does better. If points are arranged in a roughly square grid and scaled appropriately, many horizontal and vertical neighboring pairs are at distance 1. That gives about 2n unit distances, and more refined grid-based constructions improve the count slightly.
Erdős believed the best possible growth should be close to linear, up to a very slowly growing factor. The informal expectation became that u(n) should be n^(1+o(1)), meaning only slightly more than n by a sub-polynomial factor. The known upper bound has been much larger: u(n) is O(n^(4/3)), following work connected to the Szemerédi-Trotter theorem and incidence geometry.
That gap left a large unknown region. The square-grid-style lower bounds were far below the upper bound, but no one had found a construction that broke the near-linear expectation by a fixed polynomial exponent in the Euclidean plane.
OpenAI says its model has now done that. The proof constructs infinitely many point sets with at least n^(1+delta) unit distances for some fixed positive delta. In other words, the improvement is not just a small logarithmic gain. It breaks the old expectation by a genuine power of n.
OpenAI's post emphasizes the problem's place in the field. The company notes that the 2005 book Research Problems in Discrete Geometry, by Peter Brass, William Moser and János Pach, treated the question as one of the most familiar and easiest-to-state problems in combinatorial geometry. Erdős also offered a monetary prize for its resolution, a sign of how seriously he viewed the question.
What the new result claims
The official proof states that for some positive constant epsilon, there are infinitely many n such that u(n) is at least n^(1+epsilon). The proof released by OpenAI does not optimize the exponent in the main statement. OpenAI's post says a forthcoming refinement by mathematician Will Sawin obtains delta = 0.014.
That number may look small, but its mathematical meaning is large. A fixed positive exponent changes the asymptotic class of the lower bound. It says the number of unit distances can grow polynomially faster than the long-believed square-grid benchmark.
The construction is not a minor variation of the square grid. OpenAI says the model found a new family of constructions. The proof uses tools from algebraic number theory, including Gaussian integers, algebraic number fields, class field towers and ideas related to Golod-Shafarevich theory. Those are not the objects most readers would expect in a problem that begins with points and distances in the plane.
At a high level, the construction starts from algebraic structures that contain many norm-one elements. Those elements are used to build high-dimensional configurations with many unit-distance relationships. The proof then projects or represents those structures in a way that produces planar point sets while preserving enough unit distances to beat the previous lower-bound expectation.
The official material emphasizes that the proof is not a numerical experiment. It is a mathematical construction with an asymptotic lower bound. That is why the result matters: it does not only show a single surprising configuration. It gives an infinite family.
Why square grids were believed to be close to optimal
The square grid has long been the natural example for the planar unit distance problem. It is simple, visual and effective. Many points have horizontal or vertical neighbors at distance 1, and more refined versions can add additional unit-distance relationships.
For decades, improvements to the lower bound remained close to the same general family of ideas. The best known lower bounds were essentially in the n^(1+o(1)) regime, while the best upper bound remained n^(4/3). That left room for a major surprise, but no construction had shown that a fixed exponent above 1 was possible.
OpenAI's result changes that. It says the intuition that near-grid configurations define the right scale was wrong. The best examples can come from a very different mathematical source.
This is common in deep mathematics. A problem may look geometric, but the solution may come from algebra, number theory or combinatorics. The surprise here is not only that a different mathematical area enters the problem. It is that a general-purpose AI model found the route.
How the proof was checked
OpenAI said the proof was checked by external mathematicians. The company also released a companion remarks paper with contributions from Daniel Litt, Will Sawin, Victor Wang and Melanie Matchett Wood, alongside context from the OpenAI team.
The official proof includes an AI-use statement explaining how the result moved from model output to human-readable mathematics. According to that statement, the model was given an AI-written problem statement and asked to resolve the Erdős problem. The output was sent to an AI grading pipeline. After the pipeline judged the result highly likely to be correct, OpenAI researchers and mathematicians examined it. The statement says the proof was then checked through a combination of AI-assisted verification, rewriting and external human mathematical review.
That workflow is important because it separates discovery from publication. The model produced the core proof, but human experts still had to inspect, rewrite and validate it. OpenAI's announcement does not suggest that the mathematical community should accept AI output without verification. It suggests that AI may now be capable of producing the kind of original argument that human mathematicians can then check.
The companion paper also gives context about related mathematical work. It explains that variants of the unit distance problem have been studied in other settings, including non-Euclidean norms and higher-dimensional or algebraic constructions. The new proof connects the Euclidean planar problem to a richer algebraic framework.
The role of algebraic number theory
The official proof is technical, but the broad idea can be described without following every algebraic detail.
In the Euclidean plane, a unit distance can be represented using complex numbers. Points can be viewed as complex values, and a pair of points is at unit distance if the difference between them has norm 1. The Gaussian integers, which are complex numbers with integer real and imaginary parts, give a familiar arithmetic setting for such questions.
OpenAI's proof goes beyond the ordinary Gaussian integer grid. It uses algebraic number fields and towers of extensions to create many elements with controlled norms. In such settings, the arithmetic structure can produce many relationships that behave like unit distances.
The challenge is to bring those relationships back to planar geometry. A high-dimensional or algebraic object is not automatically a configuration of points in the ordinary plane. The proof has to preserve enough distance relationships while keeping the number of points under control.
The companion remarks describe the construction as drawing on infinite class field towers and bounded root discriminants, areas of algebraic number theory far from the elementary statement of the original problem. This is one reason the result is surprising: the solution path reaches into a sophisticated part of number theory to answer a geometric extremal question.
Why mathematicians are paying attention
The names attached to the official companion materials and OpenAI's announcement signal that the result is being taken seriously. The post includes comments or context involving prominent mathematicians, including Noga Alon, Tim Gowers, Arul Shankar and Jacob Tsimerman. The companion paper includes additional work from Daniel Litt, Will Sawin, Victor Wang and Melanie Matchett Wood.
The official post says Noga Alon described the work as an outstanding achievement. Tim Gowers said he expects the proof to appear in a leading journal. Arul Shankar emphasized that AI systems are moving beyond the role of helpful assistants. Jacob Tsimerman highlighted the connection between the proof and deep number-theoretic structures.
Those reactions do not replace peer review, but they matter. Discrete geometry is a field where a false proof can look convincing to non-specialists. The fact that external mathematicians examined the argument and that companion remarks were published gives the result a different status from an unverified model output.
The next step for the mathematical community will be normal scrutiny: reading the proof, checking details, simplifying arguments, optimizing constants and exploring related problems.
What this means for AI research
OpenAI is presenting the result as evidence that reasoning models can contribute to frontier knowledge, not only reproduce existing knowledge. That distinction matters.
Most AI use in mathematics has been seen in several categories: solving contest problems, suggesting lemmas, searching examples, writing code, formalizing proofs or helping humans explore a problem. Those are useful, but they are not the same as autonomously producing a proof that changes a central open problem.
If OpenAI's framing holds up under mathematical scrutiny, the result suggests general-purpose models are beginning to cross into research discovery. The model did not need a custom system built around the planar unit distance problem. It did not simply verify a known argument. It generated a new construction that experts then checked.
That does not mean mathematicians are being replaced. The official workflow still relied on human judgment at key points. Humans framed the evaluation, recognized the significance, checked the proof, rewrote the exposition and placed the result in the broader literature. The practical future is more likely to be a collaboration model: AI systems generate candidate ideas and proofs, while mathematicians test, refine and connect them to the field.
OpenAI also connected the result to the safety and alignment questions raised by stronger reasoning systems. The company argued that if models can hold together difficult arguments and produce work that survives expert scrutiny, similar capabilities could matter in biology, physics, materials science, engineering and medicine. It also said human judgment remains essential: experts still choose important problems, interpret results and decide what to pursue next.
Limits and open questions
The result is a breakthrough, but it does not solve the entire planar unit distance problem. The best known upper bound is still much larger than the new lower bound. OpenAI's proof disproves the near-linear conjectural lower-bound expectation, but it does not determine the true asymptotic growth of u(n).
The exponent also remains an area for improvement. OpenAI's announcement points to a refinement with delta = 0.014, but the gap to the n^(4/3) upper bound remains wide. Future work may improve the construction, find simpler versions, sharpen the exponent or reveal new upper-bound techniques.
There are also questions about the model and evaluation process. OpenAI has not publicly released the full model chain of thought. The company released an abridged version. It also has not disclosed all training and evaluation details behind the model. For AI research, that leaves questions about reproducibility, benchmark design and how often such discoveries can be expected.
For mathematics, the usual standard remains proof verification. The important question is not whether a model produced the argument, but whether the argument is correct. OpenAI's official materials are a strong starting point, but the broader community will still examine and discuss the proof.
Why the result is different from a contest win
AI systems have already performed well on mathematical competition benchmarks. Those achievements are impressive, but open research problems are different. A contest problem is designed to have a solution, a known answer and a bounded difficulty. A research problem may remain open for decades because the right method is unknown.
The planar unit distance problem is also central to a field. It is not an obscure puzzle chosen after the fact. Erdős's problem has been part of discrete geometry since 1946, and the gap between lower and upper bounds has been a known challenge for generations.
That is why OpenAI's announcement is receiving attention. If an AI model can find a new construction for this problem, it raises the possibility that similar systems could help in other areas where the right combination of ideas crosses fields.
The construction itself is a good example. A geometric question was attacked with algebraic number theory. Humans often make progress by importing ideas from one field into another. A general-purpose reasoning model that can search across mathematical domains could become valuable precisely because it is not locked into the most obvious approach.
What OpenAI released
OpenAI's official post links to three primary materials.
The first is the proof, titled "A disproof of the Erdős unit distance conjecture." It presents the construction and the argument for infinitely many n with more than n^(1+epsilon) unit distances.
The second is a companion paper, "Some remarks on an AI-generated proof of a theorem in discrete geometry." It provides context, explanatory remarks and perspectives from mathematicians who examined the result.
The third is an abridged chain-of-thought document. It shows selected reasoning from the model output but does not expose a full internal transcript.
Together, those materials provide the official record for the announcement. Readers who want the news-level takeaway should start with the OpenAI post. Readers who want the mathematical substance should read the proof and companion remarks.
Bottom line
OpenAI's announcement is one of the strongest public claims yet that a general-purpose AI model has contributed an original result to advanced mathematics. The company says the model disproved a longstanding belief about the planar unit distance problem, finding an infinite family of constructions that beat the square-grid benchmark by a fixed polynomial factor.
The result does not close the entire problem. It does change the lower-bound landscape and undermines a widely held expectation dating back to Erdős's 1946 question. It also provides a concrete case study for AI as a research partner rather than only a calculator, search tool or writing assistant.
The proof will now be studied by mathematicians. If it withstands broader scrutiny, the episode will be remembered not only as a discrete geometry result, but also as a marker in the development of AI systems capable of making real contributions to scientific and mathematical discovery.
FAQ
What problem did OpenAI's model address?
It addressed the planar unit distance problem, which asks how many pairs of points at distance exactly 1 can be formed among n points in the plane.
Who first posed the problem?
Paul Erdős posed the problem in 1946. It became a central question in discrete geometry.
What did the OpenAI model discover?
OpenAI says the model found a new infinite family of point configurations with at least n^(1+delta) unit-distance pairs for some positive delta, beating the long-held square-grid-style expectation.
Was the model built specifically for this problem?
OpenAI says no. The result came from a general-purpose reasoning model, not a system specifically built for math problems or for this problem in particular.
Has the proof been checked?
OpenAI says the proof was checked by external mathematicians. The company also released a proof PDF and a companion remarks paper for scrutiny.
Does this solve the entire unit distance problem?
No. It disproves a longstanding conjectural belief about the lower bound, but the exact asymptotic value of u(n) remains open.
Sources
- OpenAI official announcement: openai.com
- OpenAI proof PDF, "A disproof of the Erdős unit distance conjecture": cdn.openai.com
- Companion remarks paper: cdn.openai.com
- Abridged chain-of-thought material linked by OpenAI: cdn.openai.com
Before you move on
Global AI workflow guidance. Use this short checklist to turn the article into action.
- Check whether the tool can access private files or account data.
- Verify factual claims against primary sources before publishing.
- Keep a human review step for work that affects money, school, or customers.
This guide is written for practical user safety. For account, platform, or legal decisions, confirm critical steps with the official help center or your service provider.
