OpenAI says one of its new reasoning models has produced an original mathematical proof that disproves a famous geometry conjecture first posed by Paul Erdős in 1946, according to TechCrunch AI. The company calls it the first time AI has autonomously solved a prominent open problem at the center of a mathematical field. That’s a big claim, and OpenAI has been burned by exactly this kind of claim before.
What actually happened
For nearly 80 years, mathematicians assumed the best solutions to this problem looked roughly like square grids. OpenAI posted on X that its model found a new family of constructions that performs better, and the company published companion remarks from working mathematicians backing the result.
The endorsers aren’t lightweight:
- Noga Alon
- Melanie Wood
- Thomas Bloom, who maintains the Erdős Problems site
Bloom is the same mathematician who, seven months ago, called OpenAI’s last math victory lap “a dramatic misrepresentation.” Getting him on the record this time is the strongest signal that the proof holds up.
Why the caveat matters
TechCrunch AI reports that back in the spring, former OpenAI VP Kevin Weil claimed on X that GPT-5 had solved 10 previously unsolved Erdős problems and made progress on 11 more. It hadn’t. GPT-5 had surfaced solutions that already existed in the literature. Yann LeCun and Demis Hassabis piled on, Weil deleted the post, and the episode became a cautionary tale about AI labs declaring breakthroughs before peer review.
This time OpenAI lined up the third-party validation before posting. That’s the meaningful difference.
Why this matters beyond math
The proof came from a general-purpose reasoning model, not a system custom-built for math or for this specific conjecture. OpenAI says that’s the headline: the same model architecture that handles everyday reasoning tasks can now sustain long, intricate chains of logic and pull connections across fields that human researchers hadn’t tried.
The practical read-through:
- Research workflows shift. If general models can produce novel proofs, the line between “AI assistant” and “AI collaborator” in scientific work just moved.
- Cross-field discovery gets cheaper. OpenAI is pointing at biology, physics, engineering, and medicine as the next domains where this kind of reasoning could surface non-obvious results.
- The validation bar is now higher for everyone. After the Weil incident, no lab can announce a math or science breakthrough without independent expert sign-off attached. Expect this to become standard practice.
What stands out is how OpenAI handled the rollout. Quiet, paired with statements from named mathematicians, no founder victory lap. That’s a notable shift in tone from a company that’s been criticized for hype cycles.
What to watch next
The broader math community will now stress-test the proof in detail. If it holds, expect a wave of similar attempts from DeepMind, Anthropic, and academic labs, each trying to produce their own “first autonomous proof” in a different open problem. As Bloom put it, “AI is helping us to more fully explore the cathedral of mathematics we have built over the centuries. What other unseen wonders are waiting in the wings?”
Full details at the original TechCrunch AI report.