OpenAI Claims Unnamed Internal Model Solved 100+ Open Math Problems After One Month of Training

OpenAI says an internal model it has not named solved more than 100 long-standing mathematical problems across most areas of math after roughly one month of training, according to the company. Training reportedly began on August 28, and OpenAI claims its own mathematicians were "surprised" by how quickly the model progressed.

Among the claimed results is a solution to the Hodge conjecture, a second Millennium Prize Problem, following an earlier disputed claim that the model solved the Navier-Stokes problem. Neither the specific list of problems solved, the methodology used, nor independent verification of the results has been published. OpenAI has not disclosed the model's name, architecture, parameter count, or any benchmark scores.

The claim arrives alongside OpenAI's announcement of the Advisory Group on Mathematics and Artificial Intelligence, based at the Institute for Advanced Study, which includes Fields Medalist Timothy Gowers. The timing is notable: the announcement is a direct response to an open letter titled "A Severe Misalignment of AI in Mathematics," signed by 25 Fields Medalists, warning that AI systems optimized to solve open problems could erode conceptual understanding — which the letter's signatories argue is the actual point of mathematics.

Advisory group has no say over pace

OpenAI says the group will operate independently, can publish its own recommendations without company approval, and is unpaid. But the company explicitly carved out one area from the group's mandate: research speed. "The group will not be responsible for advising us on how to pace our internal progress on mathematics," OpenAI states. The group's role is limited to advising on how results get communicated to researchers and the public — not whether OpenAI should slow down.

A pivot from stated priorities

The announcement appears to contradict recent statements from OpenAI Chief Scientist Jakub Pachocki, who said the company had deliberately avoided optimizing for math benchmarks in favor of recursive self-improvement research. According to the reporting, OpenAI only pursued the Navier-Stokes problem after hearing rumors that a separate team, including former Anthropic researchers, had already solved it — suggesting competitive pressure, not pure research strategy, drove the effort.

Gowers: sympathetic but unconvinced

Gowers declined to sign the mathematicians' open letter, explaining on his blog that he disagrees with its framing of problem-solving as merely a means to conceptual understanding rather than a valid goal in itself. He also said his own automated theorem-proving research program at Cambridge became obsolete once large language models improved — calling it swallowing "the bitter lesson." His deeper concern is structural: if AI systems take over problem-solving, fewer people may choose to become mathematicians at all, threatening the human expertise base regardless of what AI can do. He expects LLMs capable of tackling major open problems to be publicly available "within months."

What this means

None of OpenAI's specific claims here are independently verifiable: no model name, no published proofs, no benchmark data, and no third-party review of the Navier-Stokes or Hodge conjecture results. The advisory group's exclusion from pacing decisions signals that OpenAI intends to keep controlling the speed of its math research while outsourcing only the communications strategy around it — a distinction mathematicians pushing back on the letter are likely to notice immediately.