OpenAI claims 372 math breakthroughs. No model. No prompts. Three papers pulled in 24 hours.

On October 6, OpenAI published 722 mathematics manuscripts to a GitHub repository called openai/math, each attributed to an unnamed internal model. The company says that model worked through roughly 4,000 open problems and found claimed progress on 372 of them. The headline claims include a resolution of the Unique Games Conjecture (open since 2002), a result on the 4D Kakeya conjecture, and a "quasi-Riemann hypothesis" result that some in the field are calling the most substantive approach to the Riemann hypothesis in years.

By October 7, three of those 722 papers had been withdrawn. A sign error invalidated one argument. Two dependent papers went with it. Fourteen more were revised. The repository now contains 719.

Source spread

• OpenAI — "Sharing AI progress in mathematics" — hype. Company post emphasizing Lean formalizations, IAS advisory group consultation, compute transparency, and future funding for workshops.
• Decrypt — "OpenAI's secret AI model cracked hundreds of math problems" — hype. Positive framing of scope; Abhishek Saha quoted saying the quasi-Riemann result is the most important candidate.
• Online Tech Tips — "OpenAI Math Release: What's Real" — skeptic. Notes the missing prompts, missing model name, Andrew Sutherland's unverifiability concern.
• aiweekly — papers pulled — skeptic. Covers the three-paper withdrawal and the Association for Human Mathematics response.

What's real

• Lean-verified proofs are real. A Lean formal proof is machine-checkable — it either passes or it doesn't. Anything that has a Lean verification attached is not an assertion; it's a proof in a strict formal sense. Those results are worth reading.
• The compute transparency is more than nothing. OpenAI published that the average result used roughly three hours of ChatGPT Pro-equivalent compute. That's a verifiable unit, and the estimate makes the scale legible.
• OpenAI consulted the Institute for Advanced Study's Advisory Group on Mathematics and AI on publication methodology. Real consultation, not a rubber stamp — though reasonable people can disagree about whether it was sufficient.
• Even the smaller results have value if they're real. The algebra and combinatorics progress well short of the headline claims could still be useful to the research community.

What deserves a side-eye

• The model is unnamed and unreleased. MIT's Andrew Sutherland told reporters the claims are unverifiable until the model ships. That's not cynicism. That's how reproducibility works.
• The prompts were not released. OpenAI described the process as "one prompt given to one AI agent." Until the prompts are public, that's an assertion about methodology, not methodology.
• Three papers wrong on day two. If one sign error invalidated two dependent manuscripts immediately, how many others have systematic errors waiting for someone sharp enough to spot them? The Lean-unverified results have an unknown error rate, and the company's own release note warns that "some unformalized results could have issues."
• The Association for Human Mathematics has urged colleagues to stop working with OpenAI over this release. That's an institutional math community response to what they see as a norm violation. It's not fringe.

Samwise's take

What builders need to know

• This is not a product release. The math repo is a research artifact. The model behind it is unnamed, unavailable, and probably not the thing OpenAI ships next. If you're building math-adjacent tooling — proof assistants, formal verification pipelines, theorem-proving wrappers — the event worth watching is when this model actually ships, not October 6.
• Lean formalization is the quality signal to track. The Lean-verified results are the ones safe to build on now. Everything else carries OpenAI's own "might have issues" disclaimer.
• The Association for Human Mathematics position is signal. Their call for colleagues to stop working with OpenAI is the institutional math community's response to what they see as a norm violation. If that sentiment spreads, it affects how OpenAI can collaborate on future math research — and on any future model release that requires academic validation.
• The Unique Games Conjecture result is the one to watch. If that proof holds after full community review, it matters for cryptography, approximation algorithms, and computational complexity. Worth following even if you're not in the field.

Further reading

• OpenAI — "Sharing AI progress in mathematics" — official post, with compute estimates and Lean methodology
• openai/math on GitHub — the full manuscript repository
• Decrypt — "OpenAI's secret AI model cracked hundreds of math problems" — positive framing and expert quotes
• aiweekly — three papers pulled — covers the withdrawals and the Association for Human Mathematics response
• Online Tech Tips — "What's Real" — skeptical read on methodology gaps