Your feed has probably shown you the headline: an AI just cracked a maths problem that stumped humans for 80 years. It sounds like the kind of claim that falls apart the moment you check it. This time, it did not. An AI math proof produced by an OpenAI model genuinely overturned a conjecture that Paul Erdős posed in 1946, and a second round of results in August 2026 went further still.
Here is what actually happened, what it does not mean, and why the boring technical detail — how the proof was checked — is the part that matters most.
Did AI really solve an 80-year-old math problem?
Yes. In May 2026, an internal OpenAI reasoning model produced a counterexample that disproved Erdős’s unit distance conjecture, open since 1946. Independent mathematicians reviewed the argument and confirmed it holds. It is the first time an AI system has settled a serious open question in mathematics on its own.
The result was announced on 21 May 2026 and reported by Scientific American and other outlets. OpenAI sent the working to outside experts rather than publishing a press release and hoping nobody looked closely.
Fields Medallist Timothy Gowers, who reviewed it, said that if a human had submitted the paper to the journal Annals of Mathematics, he would have recommended publication “without any hesitation”. Mathematician Daniel Litt described it as the first autonomously AI-generated result he found genuinely interesting in its own right.
What was the Erdős unit distance problem?
The question is easy to state and brutally hard to answer. Scatter a number of dots on a flat page. How many pairs of those dots can be exactly one unit apart from each other?
Line the dots up in a neat square grid and you get a lot of such pairs. Erdős suspected that grids and other highly regular arrangements were essentially as good as it gets — that no clever irregular pattern could do meaningfully better, no matter how many dots you used.
The AI found one. It built an elaborate higher-dimensional lattice and projected it down into two dimensions, producing arrangements with more unit-distance pairs than the grid, for infinitely many sizes. The technique drew on algebraic number theory, not brute-force search. A follow-up paper, Remarks on the disproof of the unit distance conjecture, is available on arXiv, and mathematician Will Sawin has since improved on the construction.
The takeaway: Erdős’s intuition was that order beats disorder in this problem, and a machine found the disordered arrangement that proves otherwise.
What OpenAI’s Astra did in August 2026
The May result could have been a one-off. On 1 August 2026, OpenAI announced that its Astra model had solved ten open problems across mathematics and theoretical computer science, spanning high-dimensional sphere packing, binary and spherical codes, arithmetic circuit complexity, operator algebras and computational complexity.
Every one of those problems had been open for at least ten years, most of them far longer, with little to no progress in between.
Two details made the announcement land harder than the first:
- The price. OpenAI put the compute cost of the successful runs at roughly $2,000 at current API rates. That figure excludes failed attempts and the staff time spent preparing and formalising the results — but it is still a startling number next to a decade of unsolved problems.
- The receipts. Each result shipped with a machine-checkable certificate published on GitHub under an open licence.
OpenAI researcher Noam Brown offered the appropriate caveat, noting there were “no Millennium Prize Problems (yet)” among them. Google DeepMind separately resolved nine further Erdős problems in the same period, so this is not one lab’s isolated fluke.
How do we know the AI math proof is actually correct?
Because a computer checked it. The proofs were formalised in Lean, a theorem prover that verifies every logical step mechanically. Anyone can download the certificate and run the check themselves. If a single step is wrong, Lean rejects it — no expertise, trust or peer review required.
This is the quiet revolution inside the story. The May result relied on a social form of validation: a group of respected mathematicians read the argument and vouched for it. That works, but it does not scale. There are only so many people qualified to referee a paper in discrete geometry, and only so many hours in their week.
A Lean certificate removes that bottleneck entirely. Verification stops being a favour you ask of busy experts and becomes a command you run. Both OpenAI and Google DeepMind now use Lean 4 as their standard verification backend for this kind of work.
The takeaway: the significant shift is not that AI can generate a hard answer cheaply, but that the answer now arrives with proof attached.
Why this is different from earlier AI maths claims
Language models have been confidently producing wrong mathematics for years. Ask a chatbot for a proof and you will often get something that reads beautifully and collapses on inspection. Plenty of earlier “AI solved X” headlines were exactly that.
Three things separate the 2026 results from the noise:
- The problems were genuinely open. Not competition puzzles with known answers hiding in the training data — questions nobody had answered.
- The verification is mechanical. A Lean-checked proof cannot be plausible-but-wrong. It is correct or it is rejected.
- Outside experts were involved before publication, not after the press cycle.
It is the same pattern seen elsewhere in AI this year: the interesting question has moved from what a model can produce to whether anyone can check it. We covered a related version of that problem in our look at what the UK safety tests really found about rogue AI agents.
Does this affect anyone outside mathematics?
Not immediately, and not in the way the headlines imply. Nobody’s job disappeared this week. But the underlying shift is worth understanding: when producing a hard answer becomes cheap, the constraint moves to proving the answer is right.
That favours fields with built-in verification. Chip design, cryptography and safety-critical software all have automated ways to check whether a proposed solution works. Those are the areas where this capability translates into practical output fastest, because a machine can grade the machine.
Fields without that safety net — strategy, medicine, law — will not see the same jump, because there is no Lean equivalent for “is this the right business decision”. If you are choosing which system to use for everyday work, our 2026 AI model comparison is a more useful starting point than any research headline.
What AI still cannot do in mathematics
Plenty. The models solved problems that were precisely stated and waiting to be attacked. They did not decide which problems were worth attacking, invent a new field, or notice that a question was badly posed in the first place.
Mathematician Melanie Matchett Wood, commenting on the May result, made the point that humans still play a vital role in validating and improving these proofs. In the August results, OpenAI staff were involved in preparing and formalising the work — the model was not left alone with a keyboard.
The honest summary: AI has become a genuinely capable research collaborator in narrow, well-defined mathematical territory. It has not become a mathematician.
Frequently Asked Questions
Which AI model solved the Erdős problem?
An internal OpenAI reasoning model produced the May 2026 disproof of the unit distance conjecture. The ten additional results announced on 1 August 2026 came from OpenAI’s Astra model. Google DeepMind resolved nine separate Erdős problems around the same time.
What is the Lean theorem prover?
Lean is software that checks mathematical proofs step by step and confirms each one follows logically. A proof written in Lean can be verified automatically by anyone with a computer, which removes the need to trust either the AI or the humans reviewing it.
Did the AI really only cost $2,000?
That figure covers the compute for the successful published runs at current API rates. It does not include failed attempts, model training, or the OpenAI staff time spent preparing and formalising the results. The real total is considerably higher.
Does this mean AI can now replace mathematicians?
No. The models solved clearly stated problems that were already known to be open. Choosing which questions matter, framing new ones and interpreting results remain human work, and researchers were involved throughout both announcements.
Can I read the proofs myself?
Yes. OpenAI published the Lean certificates for the August results on GitHub under an open licence, and the arXiv paper discussing the unit distance disproof is freely available. Reading them, however, requires substantial mathematical background.
