For 80 years, no human could crack one famous math puzzle. An AI just did it on its own, in a matter of weeks. On May 20, 2026, OpenAI said one of its AI models proved an old math guess wrong, a guess that a Hungarian math legend named Paul Erdos first made back in 1946.
Why care? It might be the first time an AI has cracked an open math problem largely on its own. Not a homework question. A brick wall that top mathematicians had been stuck on for decades.
What the problem asks
The puzzle has a fancy name, the planar unit distance problem, or Erdos problem 90. But the idea is simple. Draw some dots on a page. How many pairs of those dots can sit exactly the same distance apart? Nine dots in a three-by-three grid gives you 12 such pairs. Now scale that up to billions of dots and the question gets brutal.
Erdos guessed in 1946 that a tight, evenly spaced grid was basically the best you could do, that no clever layout of dots could beat it by more than a hair. For 80 years most mathematicians agreed with his gut, even though nobody could actually prove him right.
How the model beat it
OpenAI used a general-purpose AI, the same kind that powers a chatbot, not a special math machine. They asked it one simple thing: was Erdos right? The model spit out hundreds of pages of logic and math, then came up with a brand-new way to place the dots that beats the grid.
Mark Sellke, one of the mathematicians who checked the work, said the new layout was "totally different" from the grid idea and used extra dimensions a person would never think to try. Wait, that's the part that rattled people. The AI didn't just tweak the old answer. It found a whole new kind of answer.
Why this one is different
AI math claims have been mostly garbage this year, honestly. Back in October 2025, OpenAI VP Kevin Weil posted that GPT-5 had solved 10 unsolved Erdos problems. It hadn't. The model had just dug up answers that were already printed in old papers, and Weil deleted the post after Yann LeCun and Google DeepMind chief Demis Hassabis called him out. So the bar for believing this stuff is high; rightly so.
This time the work cleared it. Timothy Gowers, who won the Fields Medal (think of it as the Nobel Prize for math), wrote that "no previous AI-generated proof has come close" to being good enough to publish. Daniel Litt of the University of Toronto, who double-checked the work and has no ties to OpenAI, called it "the first result produced autonomously by an AI that I find interesting in itself."
TechCrunch Openai Claims It Solved Year Old
The catch worth knowing
Don't oversell it. This only works on paper, never in real life. Days after OpenAI's paper, mathematician Will Sawin pushed the result further, and his better layout only starts beating the grid at around 10 to the power of 2,000,000 dots. That's a 1 with two million zeroes after it, a number so huge it means nothing in the real world.
So nobody's building anything with this. The point isn't the dots. The point is that a normal AI model held a long, brand-new math argument together in its head and linked ideas from different parts of math, all without being told how. That same week, Google DeepMind used its own model to solve nine smaller Erdos problems.
Nature AI Cracks Year Old Mathematics Challenge
What it means for the rest of us
For now, basically nothing changes in your daily life. No app gets smarter because a chatbot beat Erdos at dot-spacing.
But the bigger signal is real, and it's the kind of thing that snowballs. An AI that can hold a long, creative train of thought across hundreds of pages is an AI that could one day help with unsolved problems in biology, physics, and engineering, where the hard part is also keeping a long argument straight. That's the bet OpenAI is making. Whether it pays off is the really interesting question, and frankly it's a much bigger one than this single proof.
