OpenAI solves the Erdős problem: an AI model just proved a geometry theorem that had resisted for 80 years

OpenAI solves the Erdős problem: an AI model just proved a geometry theorem that had resisted for 80 years

🔎 AI no longer just recites — it discovers

In May 2026, OpenAI announced that one of its internal reasoning models refuted a conjecture posed by Paul Erdős in 1946. This is not an artificially inflated synthetic benchmark. It is an original mathematical proof, validated by researchers, which uses tools from algebraic number theory to push past the known limits of the planar unit distance problem.

This result changes the status of AI in mathematics. Until now, systems like AlphaGeometry or AlphaTensor excelled on closed problems, with clear rules and a bounded search space. Here, OpenAI's model navigated an open problem — that is, a problem whose answer no one knew in advance.

The timing is not coincidental. It coincides with the rise of reasoning models (GPT-5.5, Claude Opus 4.7, Gemini 3 Pro Deep Think) that are beginning to show planning and deduction capabilities well beyond paraphrasing.

The essentials

• An OpenAI internal reasoning model has refuted Erdős's unit distance conjecture, an open problem since 1946 in discrete geometry.
• The proof relies on algebraic number theory, not computational brute-force — the model discovered a previously unknown family of point arrangements.
• This is a generalist model, not a system specialized in mathematics, which makes the result more significant for the future of research.
• OpenAI had already made a premature announcement about an Erdős problem previously — this time, the proof holds up according to the initial experts consulted.
• This result surpasses previous AI advances in mathematics (AlphaGeometry, AlphaTensor) because it tackles a truly open problem.

Recommended tools

The Erdős problem: simple to state, impossible to solve

The question is deceptively elegant. How many pairs of points can one place in a plane such that each pair is exactly at a distance of 1?

Paul Erdős formulated this problem in 1946. He conjectured that the maximum number of pairs grew in O(n^(1+ε)) — in other words, roughly linearly with the number of points.

For nearly 80 years, mathematicians have nibbled away at the bounds. The best known construction relied on square grids, which yield about n^(1 + c/log log n) pairs. Nobody had managed to do significantly better.

What makes the problem vicious is that it is open. There is no textbook answer to consult. Every improvement demands a structurally new idea on how to arrange points in the plane.

This is exactly the kind of problem where classical LLMs fail: no recognizable pattern in the training data, no known solution to reproduce.

What OpenAI's model actually found

The model didn't simply optimize an existing grid. According to details published by OpenAI and reported by The Guardian, it discovered a family of point arrangements based on algebraic number theory.

Specifically, instead of placing points on a regular square grid, the model leveraged the properties of certain algebraic number fields to generate configurations where unit distances appear more densely than in any known construction.

The proof proceeds through several steps:

Algebraic construction. The model identifies sets of points in the plane whose coordinates are elements of specific number fields. These sets have the property that many pairs are automatically at a distance of 1.

Counting. It then proves that the number of pairs at a distance of 1 in these arrangements exceeds the bound associated with the Erdős conjecture, thus refuting it.

Generalization. It shows that this family of constructions is not an isolated case but can be extended, making the refutation robust.

TechCrunch emphasizes a crucial point: this is not a system trained solely on mathematics. It is a generalist reasoning model that applied tools from different branches of mathematics to find a solution.

Why it's different from AlphaGeometry and AlphaTensor

A necessary comparison. Previous AI successes in mathematics have very different characteristics.

AlphaGeometry (DeepMind, 2024) solved olympiad geometry problems. But these problems are closed: it is known that they have a solution, the axioms are fixed, and the search space, although large, is bounded. The system combined an LLM to propose constructions and a symbolic engine to deduce.

AlphaTensor (DeepMind, 2022) found new matrix multiplication algorithms. Here again, it is a well-defined search space: we are looking for the shortest sequence of multiplications for a given format.

The Erdős problem is of a different nature. Nobody knew whether the conjecture was true or false. The space of possible proofs is not bounded by a closed set of axioms. And above all, the solution requires connecting distinct mathematical fields — discrete geometry and algebraic number theory — which demands a form of conceptual creativity.

The qualitative leap is real. Moving from optimization in a known space to discovery in an unknown space changes the very nature of what is expected of AI in science.

Mathematicians' reactions: caution and wonder

Reactions are nuanced, and that's healthy. Scientific American reports that several mathematicians are impressed by the quality of the ideas, particularly the unexpected bridge between discrete geometry and number theory.

The community's response is structured around three main points.

Wonder at the ideas. The algebraic construction is deemed "ingenious" by several researchers quoted by the New York Post. This is not a trivial trick — it is an idea no one had seen in 80 years.

Caution regarding validation. The proof has not yet gone through the full peer review process. Some mathematicians point out that proofs in number theory can contain subtle errors that only emerge after months of verification.

The ghost of the previous episode. As LODJ.ma recalls, OpenAI had already made a splashy announcement about an Erdős problem — the supposedly new solutions already existed in the literature. This embarrassing episode naturally makes the community wary.

This time, however, the initial feedback is distinctly more positive. BFM TV calls the result a "major step in mathematics," and Les Numériques highlights the independent nature of the discovery.

A generalist model, not a specialized tool

This is perhaps the most important point for the future. According to ExplainX, the model used is not a system designed exclusively for mathematics.

It is an internal reasoning model that shares its architecture with OpenAI's general models. The difference lies in chain-of-thought prompting, the ability to plan over long sequences of deduction, and probably internal verification mechanisms.

Why is this significant? Because a model specialized in discrete geometry probably wouldn't have thought to use algebraic number theory. It is precisely the fact that the model was trained on a broad spectrum of mathematics that allowed it to make this conceptual bridge.

In the current agentic leaderboard in 2026, OpenAI's GPT-5.5 dominates with a score of 98.2, followed by Gemini 3 Pro Deep Think (95.4) and Claude Opus 4.7 Adaptive (94.3). These scores reflect a reasoning capability that now goes beyond the simple reproduction of patterns.

To choose the right model based on your reasoning needs, see Google Gemini vs ChatGPT vs Claude : lequel pour quel usage ?.

What this implies for scientific research

The impact goes beyond discrete geometry. If a generalist model can refute an open conjecture in mathematics, the question becomes: what other scientific fields are vulnerable?

In theoretical physics, many open problems involve connecting different formalisms — exactly the type of leap the model made between geometry and number theory.

In theoretical computer science, conjectures about complexity (P vs NP, lower bounds) could benefit from similar approaches, even though the Erdős problem is structurally more accessible.

In computational chemistry, the discovery of new molecular configurations is formally similar to the Erdős problem: finding arrangements of elements that satisfy distance constraints.

Phys.org notes that this result shows that AI can now "autonomously tackle major open problems in mathematics" — a strong formulation that would have seemed exaggerated just a year ago.

Limitations to keep in mind

Excess enthusiasm warning. This result is major, but it does not spell the end of mathematicians.

The model did not "understand" the problem. It generated a sequence of valid deductions leading to a refutation. The distinction is important: there is no intentionality, no geometric intuition in the human sense. The model does not "see" the points in the plane.

Human verification remains essential. The proof must be read, understood, and validated by mathematicians. The model can produce proofs, but it cannot yet guarantee their correctness with absolute certainty. An elegant false positive remains possible.

One problem does not make a research program. Erdős posed hundreds of problems. Solving one, even an iconic one, does not mean that AI can systematically attack open problems. Reproducibility remains to be demonstrated.

The impact on model rankings is indirect. This result comes from an internal reasoning model that is not publicly available. Accessible models like GPT-5.4 Pro (91.8 in agentic) or Claude Sonnet 4.6 (81.4) have not yet demonstrated this level of autonomous discovery.

The link with parameter golf and model efficiency

There is an interesting irony. OpenAI did not disclose the size of the model used for this discovery. But in a context where parameter golf shows that well-trained small models often outperform poorly optimized large models, the question arises.

Does mathematical discovery capability depend on the size of the model, or on the quality of the internal reasoning? If a relatively compact model can refute Erdős, this reinforces the thesis that efficiency takes precedence over raw scale.

Self-hosted models like Kimi K2.6 (88.1 in agentic) and GLM-5 Reasoning (82) show that the ecosystem is opening up beyond OpenAI. The discovery regarding Erdős is impressive, but it could accelerate the community's conviction that giant models are not needed to conduct research.

❌ Common mistakes

Mistake 1: Confusing refutation with resolution

Saying that OpenAI "resolved" the Erdős problem is inaccurate. The model refuted the conjecture — it showed that it is false. But the original problem (finding the exact bound) remains open. The nuance is important in any article or technical discussion.

Mistake 2: Comparing with past OpenAI mistakes without nuance

Yes, OpenAI messed up on an Erdős problem before. But equating the two episodes is a mistake. This time, the proof was examined by external mathematicians and initial reactions are positive. Ignoring this difference is falling into reverse cynicism.

Mistake 3: Presenting the model as an autonomous mathematician

The model did not choose to work on Erdős. It did not decide that algebraic number theory was the right approach after a night of reflection. A human formulated the problem, configured the system, and interpreted the result. AI is a discovery tool, not an independent researcher.

Mistake 4: Deducing that all open problems will fall

The Erdős problem has properties that make it relatively "amenable" to AI: simple statement, connections with other fields, partially structured proof space. Directly transposing this result to conjectures like the Riemann hypothesis or P vs NP is a generalization error.

❓ Frequently Asked Questions

What exactly is the Erdős problem?

It is the planar unit distance problem: how many pairs of points can be placed in a plane so that each pair is exactly at a distance of 1? In 1946, Erdős conjectured that the maximum is close to linear.

Did the model work alone?

No. OpenAI researchers formulated the problem, guided the model, and verified the output. The term "autonomous" used by some media means that the model produced the proof without human intervention during the reasoning process, not that the entire project was unsupervised.

Is the proof validated?

It has been reviewed by several mathematicians with positive feedback, but has not yet gone through the full peer-review process in an academic journal. Caution is still advised.

What exact model was used?

OpenAI refers to an "internal reasoning model" without publicly naming it. It is not GPT-5.5 directly accessible via the API. It is likely a variant optimized for long-chain reasoning.

Does this call into question the work of mathematicians?

No. The model produced a proof, but it is the mathematical community that validates, interprets, and integrates it into the body of knowledge. AI becomes a powerful collaborator, not a replacement.

✅ Conclusion

A generalist model has just refuted a conjecture that generations of mathematicians had been unable to tackle in 80 years — not through brute-force, but through an original algebraic idea. This is no longer optimization in a known space; it is discovery in the unknown. The rest is just a matter of timing: when will the next open problem fall to AI?