AI Erdős problems stopped being a curiosity in 2026. On May 20, an internal OpenAI reasoning model disproved the planar unit distance conjecture — a problem Paul Erdős posed in 1946 and one he personally considered among his favorites. The model did not merely find a counterexample by search: it produced an infinite family of constructions built on deep algebraic number theory, and independent mathematicians verified the proof. Quanta Magazine's August 3, 2026 feature on why these problems keep falling to AI is the clearest sign that this has become a structural shift rather than a single result.
The reflex reaction — "AI does math now" — is both correct and useless. The interesting questions are narrower: what exactly was proved, what machinery did the model use, who checked it, and does producing a correct proof by unexplained means count as mathematical understanding?
Here is the honest breakdown, including the parts the mathematical community is actively arguing about.
Key Takeaways
- On May 20, 2026, OpenAI announced an internal reasoning model had disproved the planar unit distance conjecture, first posed by Erdős in 1946.
- The model found an infinite family of constructions using Golod–Shafarevich theory and infinite class field towers, achieving n^(1+δ) unit-distance pairs — a polynomial improvement over the square grid.
- Princeton mathematician Will Sawin later refined the constant to δ = 0.014.
- Independent verification came from named mathematicians including Noga Alon, Melanie Wood and Thomas Bloom; Fields Medalist Timothy Gowers indicated the proof would merit a top journal.
- A separate May 2026 paper resolved Erdős Problem 1196 after an amateur mathematician, Liam Price, fed the problem statement into GPT-5.4 Pro with no additional context.
What is the unit distance problem?
The setup is almost insultingly simple. Place n dots on a flat plane. How many pairs of those dots can be exactly one unit apart?
For nearly eighty years the consensus was that a square grid is essentially optimal — you cannot do meaningfully better than arranging points in a lattice. The conjecture bounded the maximum number of unit-distance pairs at barely more than linear in n.
You can see the grid behaviour with a few lines of code:
from itertools import combinations
from math import isclose, sqrt
def unit_pairs(k):
"""Count unit-distance pairs in a k x k integer grid (n = k*k points)."""
pts = [(x, y) for x in range(k) for y in range(k)]
return sum(
1 for (a, b) in combinations(pts, 2)
if isclose(sqrt((a[0]-b[0])**2 + (a[1]-b[1])**2), 1.0)
)
for k in (4, 8, 16, 32):
n = k * k
print(f"n={n:5d} unit pairs={unit_pairs(k):6d} ratio={unit_pairs(k)/n:.2f}")
Run it and you will see the ratio creep upward very slowly — which is exactly why the conjecture was believable. Grids grow the count just barely faster than linearly, and eighty years of attempts to beat them failed.
What did the AI model actually prove?
It found a completely different family of point configurations that beats the grid by a polynomial factor — n^(1+δ) unit-distance pairs for a fixed δ greater than zero — and proved the construction works for infinitely many n.
The machinery is the surprising part. According to OpenAI's announcement, the construction uses Golod–Shafarevich theory and infinite class field towers: tools from algebraic number theory that have essentially nothing to do with plane geometry on their face. Princeton's Will Sawin subsequently refined the improvement constant to δ = 0.014.
That cross-domain jump is what impressed working mathematicians more than the result itself. Finding a better arrangement of dots is a combinatorics problem; the winning move came from a branch of number theory concerned with the arithmetic of field extensions. A human might spend a career without connecting those two rooms.
| Aspect | Before May 2026 | After |
|---|---|---|
| Best known construction | Square grid, ~n^(1+c/log log n) | Infinite family, n^(1+δ) |
| Improvement constant | — | δ > 0, refined to 0.014 |
| Machinery | Combinatorial geometry | Golod–Shafarevich, class field towers |
| Status of conjecture | Open since 1946 | Disproved |
Was the proof verified by humans?
Yes, and by unusually credible ones. OpenAI published companion remarks from mathematicians who independently checked the argument. Noga Alon, a leading combinatorialist at Princeton, described the unit distance problem as one of Erdős's favorite problems. Melanie Wood and Thomas Bloom both provided supportive statements — and Bloom's involvement carries particular weight, because he had publicly dismantled an earlier, weaker AI mathematics claim.
Fields Medalist Timothy Gowers indicated he would recommend the proof for a top journal. Scientific American's coverage and Gil Kalai's blog post — Kalai being one of the field's most respected commentators on combinatorics — both treat the result as genuine.
This verification chain matters enormously. Mathematics has a hard truth condition that almost no other domain has: a proof is either correct or it is not, and enough qualified humans read this one to settle the question. That is why this result carries more evidentiary weight than any benchmark score.
This was not an isolated result
The unit distance proof is the most prominent case, not the only one. A separate May 2026 paper resolved Erdős Problem 1196, with co-authors including Terence Tao and Stanford's Jared Duker Lichtman. The origin of that result is the striking detail: Liam Price, an amateur mathematician in the US, fed the problem statement into GPT-5.4 Pro with no additional context, and the resulting insight became a published collaboration with some of the strongest analytic number theorists alive.
By July, the Xena Project blog — a hub for formalized mathematics — was posting under the title "Human mathematicians are being outcounterexampled". That is the reaction of a community adjusting its expectations in real time, not hype.
Quanta's August 3, 2026 feature makes the structural case for why Erdős problems specifically are falling: they are precisely stated, self-contained, often solvable by a single clever construction, and there are thousands of them catalogued. That is close to an ideal benchmark distribution for a search-and-verify system.
The argument the field is actually having
Physics World reported that these results have sparked genuine debate about the future of mathematics, and the disagreement is more interesting than the usual optimist-versus-doomer split.
The objection is not that the proofs are wrong. They are checkable, and they check out. The objection is about understanding. A proof produced by a process nobody can inspect, using machinery selected for reasons nobody can articulate, is correct but not explanatory. Mathematics has always valued proofs that reveal why something is true, not just that it is. A verified counterexample obtained from an opaque search is closer to an experimental result than to a theorem in the traditional sense.
The counterargument, made forcefully by several working mathematicians, is that this distinction is romantic rather than real. Much of mathematical progress has always come from constructions whose motivation was reconstructed after the fact. Erdős himself was famous for producing results by probabilistic arguments that revealed no structural reason at all.
Our read: both sides are describing the same shift and evaluating it differently. What changed is the cost of a correct answer, which has fallen sharply, while the cost of an explanation has not. Fields where verification is cheap — mathematics, formal methods, competitive programming — will feel this first and hardest.
The gap in the coverage: what makes math different from every other AI benchmark
Almost every article about this asks whether AI is now better than mathematicians. The more useful question is why mathematics produced an unambiguous result when no other domain has.
The answer is the verification asymmetry. In mathematics, checking a proof is vastly cheaper than finding one, and the check is objective. That combination is rare. In software engineering, checking that a large change is correct is often comparably hard to writing it. In medicine or law, the ground truth arrives years later, if ever.
This tells you exactly where AI-generated results should be trusted in 2026: wherever verification is cheap, independent of the generator, and binary. That is why formal proof assistants, test suites, type systems and fuzzing matter more now, not less — they are the mechanism that converts an unreliable generator into a reliable pipeline. The same reasoning underlies why agentic benchmarks like Terminal-Bench moved so sharply this year, as we covered in DeepSeek V4 Flash 0731: tasks with a hard pass/fail check improve fastest, because the model can be trained against a signal that cannot be gamed.
The corollary is uncomfortable. In domains where you cannot cheaply verify the output, none of this progress transfers. The Erdős results are not evidence that AI reasoning generalizes — they are evidence that AI reasoning plus cheap verification is extremely powerful, and that most of the world lacks the second ingredient.
Frequently Asked Questions
What Erdős problem did AI solve? An internal OpenAI reasoning model disproved the planar unit distance conjecture, posed by Paul Erdős in 1946, on May 20, 2026. Separately, a May 2026 paper resolved Erdős Problem 1196 after an amateur mathematician submitted it to GPT-5.4 Pro.
What is the unit distance problem? It asks how many pairs of points among n points in a plane can be exactly one unit apart. For 80 years the square grid was believed to be essentially optimal; the AI-found construction beats it by a polynomial factor.
Did mathematicians verify the AI's proof? Yes. Noga Alon, Melanie Wood and Thomas Bloom independently reviewed it and provided supportive statements, and Fields Medalist Timothy Gowers indicated it would merit publication in a top journal. Will Sawin later refined the improvement constant to δ = 0.014.
Which AI model disproved the conjecture? An internal OpenAI reasoning model not available to the public. The separate Erdős Problem 1196 result came from GPT-5.4 Pro, which is publicly accessible.
Does this mean AI is better at math than humans? No. It means AI can now produce publishable results on precisely-stated, self-contained problems where correctness is cheaply verifiable. Open-ended theory building, where the hard part is deciding what to prove, remains firmly human territory.
Why are Erdős problems falling to AI specifically? They are precisely stated, self-contained, frequently solvable by a single ingenious construction, and there are thousands of them catalogued with known status — close to an ideal benchmark for a system that can generate many candidates and have them checked.
The verdict
This is the most substantive AI research result of 2026, and it is substantive precisely because mathematics refuses to grade on a curve. The proof is either correct or it is not; enough qualified humans read it; it is correct. No benchmark contamination argument applies, no cherry-picked demo, no vendor-supplied evaluation.
Our verdict: treat this as a real capability milestone with a sharply bounded scope. AI systems can now contribute original, publishable mathematics on well-posed problems, and that is genuinely new. They have not demonstrated the ability to decide which problems are worth posing, which is most of what research mathematicians actually do. Expect the Erdős problem list to keep shrinking, and expect open-ended theory building to look unchanged for a while yet.
If you want the model-capability context behind results like this, our breakdowns of Claude Opus 5 and GPT-5.6 Sol, Terra and Luna cover where the current reasoning frontier sits.
Erdős used to say that a mathematician is a machine for turning coffee into theorems. Eighty years later, one of his favorite problems fell to a machine that does not drink coffee — and the proof still had to pass a human referee.
