OpenAI official logo (public domain, Wikimedia Commons) — CryptoBriefing brand treatment
An unreleased OpenAI model produced 722 manuscripts across 372 result families, and the math world is now checking the work
OpenAI has handed the mathematics community a very large stack of homework to grade. According to The Washington Post, mathematicians are now analyzing OpenAI’s findings on over 300 problems, all generated by an internal AI model the company has not released.
The scale is the story. On October 6-7, 2026, OpenAI published 722 manuscripts organized into 372 “result families,” covering number theory, algebraic geometry, analysis and theoretical computer science.
What OpenAI actually released #
The model behind the results remains internal. OpenAI put the outputs into the open, but not the system that produced them.
A key detail: many of the results come with Lean formalizations. Lean is a proof assistant, software that checks every logical step of a proof the way a compiler checks code. If a proof compiles in Lean, a machine has confirmed that each step follows from the previous one.
As of early October 2026, 42% of the results, around 300, had Lean formalizations. The model was tested against approximately 4,000 problems. The released catalog is the subset of outputs OpenAI considered significant. OpenAI also provided limited summaries to help readers make sense of a portion of the results.
Three manuscripts were retracted shortly after release due to errors. Out of 722, that is a small number, but it is also a reminder that “generated by AI” and “correct” are not synonyms.
The Millennium Prize shadow #
In September 2026, OpenAI reported progress on the Navier-Stokes equations, one of the Millennium Prize Problems. That effort reportedly involved about 10,000 AI agents working over 88 hours.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Some of the newly released manuscripts touch on other Millennium Prize Problems. OpenAI did not claim complete resolutions for any of them in this publication.
Prominent mathematicians have taken notice. Terence Tao and Alex Kontorovich are among the figures who have weighed in on the effort. Particular attention has fallen on a result connected to the Riemann hypothesis. The research does not describe the Riemann hypothesis as solved, and nothing in the release claims that it is.
The transparency problem #
Critics have raised concerns about verifiability, because OpenAI has not shared the underlying model or the complete prompts used to generate the results.
OpenAI did build in some outside structure. An advisory group from the Institute for Advanced Study, the Princeton research center long associated with pure mathematics, provided guidelines for how the project was run. Those guidelines incorporated feedback from hundreds of mathematicians and were meant to address concerns about bringing AI into traditional research.
Why this release is different #
This release is pitched as a large catalog of research-level results across multiple fields at once. The combination of volume, breadth and machine-checkable proofs is what has mathematicians paying attention, and why the research describes it as a transformative stage for the overlap between AI and math.
What this means #
For working mathematicians, the immediate job is triage. With 372 result families on the table, the community has to decide which ones deserve close human attention. The roughly 300 results with Lean formalizations are the natural starting point, since their logic has already passed a mechanical check. The things to watch are specific. Watch for further retractions, for independent verification of the Riemann hypothesis-related result, and for whether OpenAI shares more about its methods. Watch, too, for whether any of these results get cited and built upon by human researchers, which is ultimately how mathematics decides what counts.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our