What exactly are you supposed to do with a proof you can’t interrogate?
That’s the question sitting at the center of OpenAI’s October 6, 2026 release, and nearly nobody asking it is getting a straight answer. The company dropped 722 mathematical manuscripts on GitHub, credited to an internal frontier model it has not named and will not release. No press conference. No launch video. A repository.
The contents are not thin. OpenAI published 372 result families, supporting proof artifacts, Lean formalizations, and ten abridged summaries of the model’s reasoning. Lean formalizations matter here, because a machine-checkable proof is one of the few things in this story that can be verified without taking OpenAI’s word for anything. But ten abridged summaries across 722 manuscripts is a ratio that tells you something about what the company considers the product and what it considers a courtesy.
A Three-Week Setup
This didn’t come out of nowhere. In late September 2026, OpenAI announced that the same unnamed model had resolved more than 100 long-standing open problems in mathematics, including Navier-Stokes existence and smoothness. Earlier reporting from CNBC and the BBC described the Navier-Stokes result arriving after roughly 88 hours of compute, coordinating up to 10,000 AI agents.
Quanta Magazine had already been tracking the buildup through the summer of 2026, when models were offering proofs of decades-old conjectures at a pace that looked less like a trend and more like a flood. By October, Scientific American was covering the second wave as a field in shock. Wes Roth’s video on the release pulled tens of thousands of views within hours. The reaction was loud and the reaction was fast, which is usually a sign that very few people have had time to read anything.
I run reviews for a living. I look at AI tools, I test their claims, and I tell you whether they hold up. So let me be plain about the problem with this one: Neither can you. Neither can the mathematicians whose life’s work just got reorganized by a system they are not permitted to access.
What Gets Verified and What Gets Believed
There are two distinct claims tangled together here, and conflating them is how hype gets laundered into consensus.
- Claim one: these 722 manuscripts contain correct mathematics. This is checkable. Slowly, painfully, by humans and by Lean. It will take months at minimum, and the Lean formalizations give it a real head start.
- Claim two: an AI system autonomously produced this body of work. This is not checkable. Not by anyone outside OpenAI.
Claim one being true does not establish claim two. A repository of valid proofs is a repository of valid proofs. The attribution to an unreleased model is a statement about provenance, and provenance is exactly the thing a GitHub drop cannot establish. OpenAI is asking the mathematical community to accept the most consequential part of the story on faith while handing over the part that was always going to be verifiable anyway.
Mathematicians and researchers have, predictably, started pushing back on precisely this point. That is not reflexive skepticism or sour grapes. That is the discipline functioning as designed. Mathematics has spent centuries building a culture where the proof is public, the method is inspectable, and the author is available for questions. The method here is a 10,000-agent swarm nobody can run, and the author does not have a name.
Why the Delivery Method Is the Story
Releasing this on GitHub instead of through a paper, a conference, or any venue with review attached was a choice. It skips the gate. It lets the headlines form before the checking starts, and headlines formed the same day across CNBC, the BBC, Scientific American, and every AI channel with a thumbnail budget.
If the mathematics holds, OpenAI will have been right to move fast, and the ten reasoning summaries will look like generosity rather than minimalism. If significant portions of it fray under scrutiny, the company will have spent an enormous amount of the field’s attention and goodwill on work it would not let anyone examine properly.
My honest read, as someone whose job is calling this stuff: the Lean formalizations are the only reason to take this seriously rather than treat it as a marketing event, and the withheld model is the only reason not to take it completely seriously. Both things are true at once, which is deeply unsatisfying and also the actual state of the evidence.
The useful posture right now is patience. Let the checkers check. Watch what survives the Lean pipeline and what gets quietly revised. Pay attention to which mathematicians say the proofs taught them something versus which say they merely confirmed something. That distinction will tell you more about whether a machine did mathematics than any announcement will.
Until the model is accessible, 722 manuscripts is a very large pile of homework with no one in the room to answer for it.
🕒 Published: