OpenAI vice president Kevin Weil said in a post on X that GPT-5 had found solutions to ten previously unsolved Erdős problems, with progress on eleven more, describing them as open for decades. Other OpenAI researchers amplified the claim. The post has since been deleted.

Thomas Bloom, who maintains the problem list at erdosproblems.com, pushed back within hours. On his site, “open” means he personally is unaware of a paper that solves the problem, not that no solution exists. What GPT-5 produced, Bloom said, were references to existing literature he had missed.

Google DeepMind chief executive Demis Hassabis called the episode embarrassing. Meta’s Yann LeCun said OpenAI had believed its own hype. Sébastien Bubeck, the OpenAI researcher who had been touting the results, conceded that only solutions already in the literature were found, while arguing that tracking down a buried reference is itself difficult work.

The exchange becomes a story about a word doing too much work. “Open” on a hand-curated list is not a proof of absence, and a model that reads the literature is not a model that writes new mathematics. The more interesting claim — that GPT-5 is good at finding papers — gets buried under the retraction.