OpenAI's Internal Model Cracked 100+ Open Math Problems in a Month, Company Says

OpenAI says an internal model solved more than 100 open mathematics problems after roughly a month of training, a claim that lands as the company faces mounting scrutiny from working mathematicians over how it presents such results. The model has not been released publicly, and OpenAI has not published the full list of problems or the proofs it says were generated. The disclosure was reported by The Decoder.

The company is pairing the claim with a financial commitment to an independent advisory group based at the Institute for Advanced Study, the Princeton, New Jersey institution long associated with figures such as Albert Einstein and Kurt Gödel. According to the report, that group will operate independently, but OpenAI has explicitly kept its own research pace outside the scope of the group's advisory role. That carve-out matters: it means the external body can comment on mathematical questions without having a mandate to weigh in on how quickly OpenAI pushes out new models.

The timing is not incidental. Mathematicians have spent the past year picking apart AI-generated proofs and pushing back on announcements that treat benchmark wins as mathematical discovery. Verifying a proof is slow, specialized work, and a model that produces a plausible-looking argument can still be wrong in ways only an expert referee would catch. Without the problem list, the proofs, or independent verification, the 100-plus figure is a company claim rather than a replicated result.

The structure of the arrangement is itself the story. An advisory group funded by the company whose work it might critique, and whose remit excludes the company's research speed, is a narrower form of oversight than the phrase 'independent advisory group' might suggest. It echoes a pattern across the industry, where labs sponsor external ethics or safety bodies while retaining control over what those bodies can examine. Whether the IAS group sets its own agenda or works from questions OpenAI finds useful will determine how much weight its output carries.

What to watch next is concrete. OpenAI could publish the problems, the proofs, and the training details, which would let mathematicians check the claims directly instead of taking them on faith. The advisory group could also say publicly what it will and will not review, and who picked its members. Until one of those happens, the 100-problem claim sits in the same category as other unreleased-model announcements: notable, unverified, and dependent on the company's own account.