Outdated: GPT-5 only found old Erdős answers (2025). AI has since solved 17 of 68 major ones
In 2025 an OpenAI executive said GPT-5 solved 10 open Erdős problems; it had found old answers. Since 2026, AI solves open ones itself.
This is a dated record. In October 2025, OpenAI VP Kevin Weil posted that GPT-5 had found solutions to 10 previously unsolved Erdős problems (maths questions posed by Paul Erdős). Thomas Bloom, who runs erdosproblems.com, pointed out that “open” on his site only meant he knew of no published solution. GPT-5 had found papers that already solved them. Weil deleted the post.
Within months that was no longer the state of play. In January 2026, Problem #728 became the first widely regarded as solved by AI on its own (GPT-5.2 Pro, with a Lean proof from Harmonic’s Aristotle). By June 2026, Terence Tao’s community wiki listed 19 full solutions by AI alone and 37 more by AI working with people. In May 2026 an OpenAI model disproved Erdős’s 80-year-old unit distance conjecture, and nine mathematicians verified it.
On Epoch AI’s benchmark of 68 significant open Erdős problems, 17 have now been solved by AI. Under Epoch’s strict rules (one attempt, a $300 budget, a computer-checked proof) the best model, GPT-6.1 Sol, solves 3.
What it means
Finding an answer that already exists in old papers is useful, but it is not solving a problem, and the 2025 claim mixed the two up. Since then AI has produced genuinely new solutions. It now solves some major open problems, though only 3 of the 68 under strict test conditions.
Caveats
- The October 2025 claim was wrong when it was made. This entry is kept as a record of that claim, not as a measure of what AI can do now.
- Easy or little-studied problems are over-represented among AI successes, and some “AI solutions” later turned out to be in the literature.
- 12 of the 17 come from OpenAI’s 6 October 2026 research release, outside Epoch’s official runs.
- Weil’s original post was deleted; its wording comes from TechCrunch’s report.