Remember September? OpenAI said its system had worked through the Navier-Stokes existence and smoothness problem, reportedly chewing roughly 88 hours of compute across as many as 10,000 coordinating agents. The claim bounced from CNBC to the BBC inside a day. I remember reading it and thinking the same thing I think about most launches that cross my desk: fine, but where do I sign in?
That question has aged badly. On October 6, 2026, OpenAI dropped 722 mathematical manuscripts on GitHub, crediting the work to an internal frontier model it hasn’t named and won’t release. Not a keynote. Not a demo. A repository. The bundle includes 372 result families, supporting proof artifacts, Lean formalizations, and ten abridged summaries of the model’s reasoning. That follows the late-September announcement that the same model had resolved more than 100 long-standing open problems.
What I review, and why this breaks my process
My job here is simple. Something ships, I run it, I tell you whether it earns a slot in your stack. I poke at the edges, I try to make it fail, and I write down what happens. That process needs one ingredient: access.
There is no access here. The model is internal. The output is public. Those two sentences together describe something I genuinely cannot evaluate as a tool, because it isn’t one yet. It’s a body of claimed results with a byline nobody outside the company can inspect.
Scientific American covered it the same day. Quanta has been tracking the broader pattern since the summer, when proofs of decades-old conjectures started appearing week after week. Wes Roth put up a video that cleared 34,000 views before the day was out. The attention is real. The reproducibility is a different story.
The Lean files are the interesting part
Here’s where I’ll give credit. Including Lean formalizations was the right call, and it’s the one piece of this release that does something concrete for skeptics. Lean checks proofs mechanically. If a formalization compiles, a machine has agreed the argument holds. You don’t need to trust OpenAI’s characterization of its own model to run a proof checker.
So the release splits cleanly into two categories, and anyone writing about this should keep them apart:
- Machine-checkable claims. The Lean formalizations. These can be verified by anyone with the toolchain, independent of who or what produced them.
- Unverifiable claims. The attribution. That an unnamed internal model generated this work, at the scale described, with the reasoning summarized in those ten abridged write-ups.
The first category is evidence. The second is a press statement. Mathematicians and researchers have been questioning exactly this gap since the drop, and they’re right to. A correct proof tells you the math is correct. It tells you nothing about the provenance.
Ten summaries for 372 result families
Ten abridged summaries of the model’s reasoning, covering 372 result families. Do that division and you get a sense of how much of the process stayed behind the curtain. I’m not accusing anyone of anything. I’m pointing out that if your central claim is about how the results were produced, and you ship summaries for a small slice of them, you’ve chosen to make the central claim the hardest part to check.
Compare it to how I’d handle any other product. If a vendor told me their tool automated 372 workflows and sent me ten case studies plus a pile of output files, I’d call that a portfolio, not a benchmark. Good portfolio. Still a portfolio.
What this means for anyone building with AI
Practical read for the people who use this site: nothing in your stack changes this week. There’s no API, no weights, no pricing page. If you work in formal verification, the Lean artifacts are worth your time on their own merits. Everyone else is watching a research announcement, and watching is all that’s on offer.
The longer-term signal matters more. We’re moving toward a world where frontier capability gets demonstrated through artifacts rather than access, and that shifts how evaluation has to work. Reviewers like me lose the ability to test. What’s left is checking outputs with independent tools, which is exactly why those Lean files carry more weight than the rest of the package combined.
My honest position: the math may well hold up, and proof checkers will settle that part. The story about the model remains a company telling us about a company’s model. Those are separate questions, and the GitHub drop answers one of them.
I’d love to test the thing. Send me a login and I’ll write the real review.
🕒 Published: