722 Papers in One Night: Did AI Just Crack 377 Unsolved Math Problems?
October 8, 2026 · 5 min read
On the evening of October 6, OpenAI dumped 722 manuscripts onto GitHub in one go. The claim attached: an unreleased frontier model has solved or advanced 377 mathematical problems — spanning algebra, number theory, topology, and computer science — including work touching the Riemann Hypothesis and the Birch–Swinnerton-Dyer formula, both Millennium Prize problems. OpenAI says each result cost roughly three hours of ChatGPT Pro-level reasoning compute on average, and that every proof has been formalized in Lean, a machine-checkable proof language. None of it has been peer-reviewed. The math world is in an uproar: a researcher quoted by The Wall Street Journal called it "the most significant moment in the history of mathematics." (via Ynetnews)
If that sounds too big to be real, you're thinking clearly. This is the kind of announcement that splits the room: one half sees a genuine turning point, the other sees a very large press release. Both halves have a case — and both deserve to be heard, because the details cut in more than one direction. So let's take the claims one at a time, in the open, and give excitement and skepticism equal airtime.
377 problems: what does that number even feel like?
A working mathematician might prove one or two substantial theorems in a good year; a single hard problem can eat a decade. Now picture 377 such results arriving in one upload, each supposedly generated in about three hours of compute. That is not an incremental step — it is a different production function for mathematics. Even if only a quarter of these survive peer review, it still rewrites the pace of the field, and the reviewers' workload would dwarf anything the journals have handled before. And if most don't survive, the number itself was doing most of the arguing: 377 is a headline figure, and headlines don't get peer-reviewed.
Lean formalization: the machine-checkable part
The most important word in OpenAI's announcement may be "Lean." A Lean proof is not a persuasive essay — it is a program a computer can check line by line. Ordinary math papers can hide gaps that the community only finds years later, when somebody finally spots the crack. A formalized proof in Lean is verifiable by anyone with a laptop. That doesn't make the claims true, but it makes them checkable — which is exactly what peer review needs most. If these 722 manuscripts really are Lean-formalized, reviewers can skip the "did the algebra go wrong" stage and go straight to the ideas: are the conjectures interesting, the methods new, the theorems worth their weight? Critics have pointed out that formalization can also check a perfectly rigorous version of a weak result — a correct proof of a marginal claim still isn't a breakthrough. True. But it does move the argument from the basement to the ground floor: the conversation shifts from "is this right" to "does this matter," which is a much better fight to have.
"Not yet peer-reviewed": who exactly is losing it?
Read the fine print on every report: none of these papers has been peer-reviewed. The Wall Street Journal quote — "the most significant moment in the history of mathematics" — came from a researcher, but right now the explosion is mostly media-shaped. Mathematicians will do what they always do: check the proofs, reproduce them, and argue about them for months — the process moves at the speed of careful people, not at the speed of press cycles. The excitement is real, but so is the long history of AI breakthrough claims that shrank under scrutiny. And notice what OpenAI's numbers actually promise: "solved or advanced." Advanced is doing a lot of quiet work in that phrase. Hold both thoughts at once.
AI and mathematicians: rivals, or a new division of labor?
If machines can generate candidate proofs at this scale, the mathematician's job changes shape: less time grinding through lemmas, more time choosing which mountains are worth climbing. The people who asked the right questions, built the formalizers, and will referee the results are still human. AI doesn't make mathematicians obsolete — it moves them up the stack, from proving to judging. The romantic version of mathematics was never about churning through algebra by hand; it was about knowing where the interesting questions live. That part hasn't been automated, and this week's news doesn't change it.