- Math Release: OpenAI has published 722 manuscripts grouped into 372 result families, giving mathematicians new AI-generated work to examine.
- Proof Checking: Some results include Lean formalizations, allowing computers to check the logic of precisely stated proofs.
- Model Access: The internal model remains unreleased, and OpenAI has published selected summaries of the model’s reasoning.
- Field Response: Mathematicians differ over the release’s value, balancing access to answers against concerns about reproducibility, explanation and attribution.
OpenAI has released hundreds of proposed mathematical results, giving researchers new work to examine across mathematics and theoretical computer science. One claimed advance concerns the Riemann hypothesis, a major question tied to the distribution of prime numbers.
The collection contains 722 manuscripts organized into 372 families of related results. A family can combine a principal result with companion arguments, consequences or alternative proofs. The papers come from an internal model that OpenAI has yet to release, and the collection includes work at different stages of verification.
What the Riemann-Related Result Claims
The Riemann zeta function is a mathematical object whose behavior is closely connected to the distribution of primes. Its zeros are inputs at which the function equals zero. Mathematicians plot these inputs on a plane with real and imaginary coordinates.
OpenAI’s stated formalization describes a zero-free half-plane for the zeta function: no zeros where the input’s real coordinate is greater than 7/8. That is a narrower statement than the full Riemann hypothesis, which places all nontrivial zeros on a particular vertical line.
Rutgers mathematician Alex Kontorovich greeted the Riemann-related claim enthusiastically, comparing its significance to work that could immediately earn a Fields Medal, a major mathematics award.
Quasi-RH?!?!???! Are you kidding me? If a human did this, it would be an instant Fields Medal, no questions asked.
RH says zeta has no zeros in Re(s)>1/2. The best we had until a second ago was a region that got thinner and thinner the higher up the imaginary axis you go. I… https://t.co/x28qPkMP9r pic.twitter.com/EtF9mGTpum
— Alex Kontorovich (@AlexKontorovich) October 6, 2026
What a Computer-Checked Proof Establishes
Lean expresses mathematical definitions and proof steps in a form that software can check. With a trusted checking setup, the Comparator tool tests whether a submitted proof establishes the target formal statement using only permitted assumptions. Mathematicians still need to ensure that the formal statement captures the question they intended to ask.
OpenAI’s formalization catalogue lists material associated with 162 of the collection’s 722 manuscripts. The company warns that some unformalized results could have issues and says it will add formalizations and preserve corrections as new versions.
Why the Generation Claims Remain Contested
An OpenAI spokesperson told Scientific American that almost every new result came from a single prompt handed to one AI agent, while acknowledging that some results might have taken multiple attempts. The report contrasts that arrangement with the company’s earlier Navier-Stokes work, concerning fluid-flow equations, which involved a 10,000-agent swarm and millions of dollars in computing power.
OpenAI says it posed approximately 4,000 problems to the internal model. It reports an average computational effort per result equivalent to three hours of ChatGPT Pro thinking compute. The Riemann-related work is among the exceptions to the company’s usual procedure.
The public disclosure includes ten abridged reasoning summaries. Exact prompts and per-result attempt counts remain undisclosed, leaving outsiders without the information needed to repeat the reported production process.
MIT mathematician Andrew Sutherland told the magazine that claims about solving problems with one prompt and one agent should remain unverified until the model is released and others can replicate the results. “We should ask for receipts.”
Answers, Understanding and the Release Debate
University of Toronto mathematician Daniel Litt sees value in making the answers available rather than keeping them secret. “To me, it’s going to be a good thing for mathematics.”
In a declaration predating the release, Fields Medalists including Terence Tao and Peter Scholze argue that mathematics aims to build conceptual understanding, with solved problems serving that larger purpose. Their concern centers on what follows a proof: talks, discussion, simplification and careful writing make new methods understandable and reusable. Rushed announcements can leave ideas unexplained and relevant prior work uncredited, weakening the human process that brings results into the mathematical canon.
After the controversy over OpenAI’s September Navier-Stokes release, OpenAI announced an independent mathematics advisory group on September 21. Scientific American reports that the group’s recommendations call for disclosure of the model, exact prompts and compute time behind each result.
OpenAI told the magazine that it was taking the guidance seriously and trying to comply. The company also said it was working to release the model responsibly and defended continued testing on open mathematics problems as necessary to measure improving AI.
Its spokesperson acknowledged that many of the newly released results were not yet understood by OpenAI’s own mathematicians.
