OpenAI Claimed Its AI Solved 100+ Open Math Problems. Here’s What We Know.

OpenAI’s September 21 announcement was an extraordinary research claim, not a settled mathematical result. The company said that an unnamed internal model, whose training began on August 28, had resolved more than 100 long-standing open problems across most areas of mathematics. It made the statement while announcing an independent mathematics advisory group hosted by the Institute for Advanced Study. 515

The crucial distinction is between a company’s claim of a solution and a solution that mathematicians have independently reviewed and accepted. Publicly available material supports the former; it does not yet establish the latter for the full set of claimed problems.

What OpenAI said it had achieved

The September announcement followed OpenAI’s September 8 claim that roughly 10,000 AI agents, using an unreleased internal model, found a finite-time singularity for forced three-dimensional Navier–Stokes equations in 88 hours. OpenAI described the result as a solution to the Navier–Stokes existence-and-smoothness problem and said it had produced both an analytical proof and a Lean formalization. 113

On September 21, the company broadened that claim: beyond Navier–Stokes, it said the internal model had resolved more than 100 long-standing open problems. 515

That wording matters. “Resolved” is OpenAI’s characterization, not an indication that each result has been published, checked by independent experts, or accepted by the mathematical community.

What evidence was released—and what was not

For the Navier–Stokes result, OpenAI publicly described an analytical proof and a formalization in Lean, a proof-assistant language. A Lean-checked result is meaningful evidence: it can establish that a precisely encoded theorem follows from the encoded definitions, axioms, and prior formal library. 13

But the public record described for the broader “100+” claim is far thinner. The September 21 announcement did not supply a public, itemized catalogue of the problems, their exact mathematical statements, proofs, formalization status, provenance, or independent-review status. Without that record, outsiders cannot audit the headline number problem by problem. 711

Nor did the available reporting establish a completed independent human peer review or general mathematical acceptance of the Navier–Stokes proof. Those are different standards from making a proof artifact available or having code compile. 146

Is the Hodge conjecture solved?

No public evidence in the supplied record supports treating the Hodge conjecture as solved.

The Hodge conjecture is one of the Clay Mathematics Institute’s Millennium Prize Problems. 37 Reporting indicated that OpenAI had acknowledged substantial progress on a second, unnamed Millennium problem, but it had not publicly identified that problem as the Hodge conjecture or released a proof for it. 14

A report, rumor, or claim of progress is not a substitute for a published proof that the research community can examine. Until a precise statement and supporting work are public and independently assessed, the responsible conclusion is that an OpenAI solution to the Hodge conjecture has not been established.

Why a Lean proof can be valid but still hard to evaluate

Formal verification and human mathematical understanding solve related, but different, problems.

A Lean formalization can provide strong assurance that an encoded theorem has no logical gap relative to its formal setup. Yet formal proof scripts often make routine mathematical reasoning explicit through many definitions, lemmas, library dependencies, rewrites, and technical steps. That can make a proof difficult for experts to read as a conceptual argument.

This is not an argument that “it compiles” is worthless. It is an argument that compilation alone does not answer every question mathematicians need to answer:

• Does the formal statement precisely match the intended Millennium problem?
• What is the core idea behind the construction or singularity mechanism?
• Which estimates or arguments do the real intellectual work?
• Is the result robust, novel, and presented in a form other researchers can build on?
• How should credit be assigned for related prior ideas?

Formal checking can be a major component of verification. It does not, by itself, create a concise human exposition, settle provenance, or produce community consensus.

The backlash: credit, priority, and research norms

The announcement generated criticism beyond the technical merits of the proof. Twenty-five Fields Medal recipients signed an open letter, A Severe Misalignment of AI in Mathematics, objecting to an AI-lab race to use famous mathematical problems as capability demonstrations. Reporting on the letter said the signatories warned that such incentives could harm mathematics, including its research culture and training. 1921

A separate priority dispute involved mathematicians Tristan Buckmaster and Levent Alpöge. In a contemporaneous statement, Buckmaster said he had been told that an OpenAI internal system had a proof of finite-time blow-up for forced Navier–Stokes equations. He also said that he and Alpöge felt pressure to release related work before their verification and exposition were complete; his statement notes that a separate hypo-dissipative Navier–Stokes result was not released because its Lean verification had not finished. 17

Buckmaster’s account is important evidence of a dispute over timing, credit, and publication pressure. It is not, on its own, proof that OpenAI copied anyone’s work. OpenAI said it began its work after hearing of a rumored breakthrough by the two mathematicians and had offered a concurrent release that would recognize their priority. 32

The unresolved issue is therefore not simply whether an AI system can generate a proof. It is also whether AI-assisted discoveries can be released in a way that permits careful attribution, independent review, and normal scholarly exchange.

What the IAS-hosted advisory group can—and cannot—do

OpenAI announced an unpaid, independent nine-member Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study in Princeton. Reporting identified members including Timothy Gowers and Martin Hairer. 45

Its stated purpose is to give mathematicians a role in evaluating and communicating AI-generated mathematical work, including advice on review and release practices. 515

However, the group is advisory rather than regulatory. Available reporting indicates it cannot compel OpenAI to release underlying proofs, dictate the company’s research schedule, veto announcements, or conclusively decide authorship. 7

That limitation is central: expert advice can improve transparency and communication, but it cannot by itself turn an internal company claim into a community-accepted theorem.

Why no Millennium Prize is available yet

Clay’s rules are explicit. Before the Clay Mathematics Institute will consider a proposed Millennium Prize solution, it must be published in a qualifying outlet, at least two years must pass after publication, and the work must receive general acceptance in the global mathematics community. 46

A company blog post, preprint, code repository, or advisory-group assessment does not automatically meet those conditions. The Clay Mathematics Institute also noted after the Navier–Stokes announcement that its rules govern both evaluation and the assignment of credit. 42

So, even if OpenAI’s Navier–Stokes result or any later result ultimately proves correct, the September announcements did not make them immediately prize-eligible. The work would still need to enter the established publication and acceptance process.

Bottom line

OpenAI claimed that an internal AI model had solved more than 100 open mathematics problems, following its earlier claim of a Lean-formalized Navier–Stokes solution. The Navier–Stokes work has public artifacts; the wider portfolio does not yet have a public, problem-by-problem evidentiary record. 1311

For now, the appropriate verdict is neither dismissal nor acceptance: these are potentially consequential claims that require transparent proofs, expert scrutiny, careful attribution, and time. That is especially true for any suggestion that another Millennium problem, such as the Hodge conjecture, has been solved.