OpenAI did not announce a proof of the Riemann hypothesis. It released a large collection of AI-generated mathematical manuscripts, including work that it says makes progress toward famous open problems. That difference matters: progress on a problem is not the same thing as solving it, and a manuscript is not automatically an accepted result.
The October 6 release is still unusually ambitious. OpenAI published hundreds of results at once, paired many of them with formal verification files, and invited mathematicians to inspect the work in public. Here is what the release actually contains and how to read the boldest claims.
What OpenAI published
According to OpenAI’s official announcement, an internal frontier model produced results across many areas of mathematics. The company shared reasoning summaries, information about computing time, and a public repository so researchers can examine the claims.
A report from Engadget counts 722 manuscripts covering progress or solutions related to 372 problems. The collection is not one giant proof. It is a batch of separate claims with different levels of importance, difficulty, and supporting evidence.
OpenAI says the average result used computing comparable to roughly three hours of extended reasoning in ChatGPT Pro. The research model itself has not been released. That limits independent reproduction for now, even though the manuscripts and many verification artifacts are public.
No, the Riemann hypothesis is not solved
The Riemann hypothesis is one of the seven Millennium Prize Problems and concerns the distribution of prime numbers through the zeros of the Riemann zeta function. A valid proof would be a historic result, reviewed intensely by specialists and eventually accepted by the mathematical community.
OpenAI’s release describes progress toward problems connected to the Riemann hypothesis. It does not claim a complete proof of the hypothesis itself. Headlines that collapse “progress toward” into “solved” erase the central fact readers need.
The same caution applies throughout the collection. Some entries may resolve precisely defined subproblems. Others may extend an existing method, tighten a bound, or offer a new route that specialists still need to test. Importance cannot be measured by manuscript count alone.
What Lean verification does and does not prove
OpenAI included Lean formalizations for many results and says more are coming. Lean is a proof assistant: a human or machine expresses a theorem and its proof in a strict formal language, and the software checks whether each logical step follows from the rules.
That is stronger than simply asking another language model whether an argument looks correct. A successful Lean check can catch hidden logical gaps and ambiguous steps. It also makes the exact assumptions easier to inspect.
Formal verification does not settle every research question. Reviewers still need to confirm that the formal statement matches the intended mathematical claim, that the assumptions are appropriate, and that the result is genuinely new. Lean can validate a carefully encoded proof without deciding whether the theorem is important or whether prior literature already contains it.
Why human review is the real bottleneck
Publishing 722 manuscripts creates a verification problem of its own. Specialists must trace citations, compare claims with existing work, check formalizations, and determine which results deserve attention. A model can generate candidates faster than a community can responsibly absorb them.
OpenAI says it consulted an advisory group and is funding workshops to support scrutiny. The public repository should allow corrections and discussion, but acceptance will happen result by result, not through one blanket verdict.
This is similar to the broader shift toward persistent AI systems discussed in our guide to OpenAI Dots and always-on agents. More autonomy increases output, but it also raises the value of reliable checks. Our comparison of OpenAI Dots versus ChatGPT explains why a specialized system should not be judged like a general chatbot.
The practical takeaway
The release is evidence that AI systems can produce mathematical research candidates at an unprecedented scale. It is not evidence that every candidate is correct, novel, or consequential. The strongest parts of the project are its public artifacts and formal checks, because they give experts something concrete to audit.
For now, the accurate summary is simple: OpenAI published a very large research package, some claims may prove valuable, and the Riemann hypothesis remains unsolved unless and until a complete proof survives expert review.