An infographic listing the 7 Millennium Prize Problems has been making the rounds again — seven $1-million questions the Clay Mathematics Institute selected in 2000, one line each: P vs NP, the Riemann Hypothesis, Yang-Mills, Navier-Stokes, Hodge, Birch and Swinnerton-Dyer, and Poincaré. It's a good list. It's also, right now, an unusually good test of how much AI can actually do in mathematics — because two of these seven problems have been at the center of real AI news in the last six weeks, and one of those stories is a cautionary tale about believing the headline before checking the proof.
Here's the honest scorecard, problem by problem, as of September 2026: what's genuinely been done, what's been claimed, and where the claims and the reality diverge.
TL;DR — the 7 problems and AI's real track record
| Problem | What it asks | AI progress as of Sept 2026 | Status |
|---|---|---|---|
| P vs NP | Is verifying a solution as easy as finding one? | None — no AI system has produced progress on the core question | Unsolved |
| Riemann Hypothesis | Do all non-trivial zeta zeros lie on one line? | Real, narrow: Claude pushed a related lower bound from 41.6% to 67.2% — not progress toward the hypothesis itself | Unsolved |
| Yang-Mills & Mass Gap | Do particles get mass from a positive mass gap in quantum field theory? | None publicly reported | Unsolved |
| Navier-Stokes Existence & Smoothness | Do 3D fluid equations always stay smooth? | Claimed, contested: OpenAI's agent swarm solved a forced variant, not the prize's natural case — see below | Unsolved (prize not claimed) |
| Hodge Conjecture | Which shapes in higher dimensions are describable algebraically? | None publicly reported | Unsolved |
| Birch and Swinnerton-Dyer | How many rational points does an elliptic curve have? | None publicly reported | Unsolved |
| Poincaré Conjecture | What 3D shapes are topologically spheres? | None — solved by a human, Grigori Perelman, in 2002-2003 | Solved |
Six unsolved, one solved — by a person, before generative AI existed. That's the frame everything below sits inside.
1. P vs NP — the one AI can't even start on
The question: if you can quickly check a solution to a problem, can you also quickly find one? Formally, does P (problems solvable quickly) equal NP (problems checkable quickly)? Most computer scientists believe the answer is no — but nobody has proven it. A proof either way would reshape cryptography, optimization, and how we think about the limits of computation.
There is no AI progress to report here, and there's a structural reason why. Every genuine AI math result so far — the Riemann bound, the Fermat formalization, even OpenAI's contested Navier-Stokes claim — worked by recombining existing human techniques to extend a known result incrementally. P vs NP doesn't offer that path. There's no partial numerical bound to nudge, no known "almost there" proof to complete. It's a binary structural question that has resisted every proof technique mathematicians have tried for over 50 years, and nothing about how current AI systems search — wide exploration of variations on known methods — is suited to inventing a genuinely new proof technique from nothing.
Where AI does intersect with P vs NP-adjacent territory is applied: cryptography built on the presumed hardness of NP problems (like RSA factoring) is something AI systems can attack computationally, even without touching the underlying theorem. explainx.ai covered a case of exactly that kind of applied progress — RSA-260 getting factored, with the work published so you can verify it yourself — which is a reminder that "hard problem gets attacked" and "hard problem gets proven" are very different events, and headlines routinely blur them.
2. The Riemann Hypothesis — where AI has its one real, verified win
The question: the Riemann zeta function's non-trivial zeros are conjectured to all lie on a single vertical line — the "critical line." That would explain deep structure in how prime numbers are distributed. Proven since 1859 to be true for some zeros, never proven for all.
This is the one problem on this list where AI has done something real and independently checkable. In August 2026, Anthropic reported that an unreleased research version of Claude — orchestrating roughly 60 subagents across two sessions and about 31 million output tokens — raised the proven lower bound on the fraction of zeros known to sit on the critical line from 41.6% to 67.2%. That's the first time that bound has cleared 50%, and it came with a machine-checkable Lean formalization plus review from two Anthropic mathematicians and two outside experts, Brian Conrey and Dan Goldston — the latter pair having personally authored parts of the prior-art machinery Claude's argument builds on.
Be precise about the size of this, because it's easy to overstate. Anthropic itself says: "we don't expect that the techniques Claude used will lead to proving the Riemann hypothesis." Claude recombined two existing lines of published human research (work by Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh, plus a 2000 Bombieri paper) into a sharper bound — a genuinely useful extension of human mathematics, not a step toward the hypothesis itself. Proving 67.2% of zeros lie on the line and proving 100% of them do are different orders of problem entirely.
3. Yang-Mills Existence and Mass Gap — silence
The question: quantum field theory predicts that particles like protons and neutrons get mass partly from the strong force described by Yang-Mills theory — but nobody has rigorously proven that the theory has a "mass gap" (a minimum positive energy above the vacuum state) using accepted axioms of quantum field theory. Physicists use Yang-Mills successfully every day; mathematicians can't yet prove why it's mathematically consistent.
There is no publicly reported AI attempt on this problem, contested or otherwise, as of September 2026. It sits in a category the AI math results above haven't touched: it requires formalizing physical intuition into rigorous mathematical axioms across quantum field theory, a domain where even stating the target theorem in a form a proof assistant like Lean could check is itself an unsolved research problem. None of the labs that produced this year's other math results — Anthropic on Riemann and Fermat, OpenAI on Navier-Stokes and the Erdős planar unit-distance problem — have announced work here.
4. Navier-Stokes Existence and Smoothness — the cautionary tale
The question: do the equations describing fluid motion in three dimensions always produce smooth, well-behaved solutions for all time — or can a solution "blow up," developing a singularity in finite time despite viscosity's smoothing effect? This is the problem the infographic flags as having a "recent AI claim, not yet verified," and it's worth walking through carefully, because it's the clearest example on this list of the gap between a viral headline and what actually held up.
The rumor first. In early September 2026, a claim spread across X that Claude had solved Navier-Stokes. explainx.ai fact-checked it the same week: the entire claim traced to one X user's explicitly labeled prediction — not a report of anything Anthropic had announced, confirmed, or submitted anywhere. No Anthropic source, no paper, no Clay Institute submission. It gained traction largely because Anthropic's real Riemann result (above) had just made "Claude does something legitimate to a Millennium Prize problem" feel plausible — which is exactly the mechanism that makes a false claim spread faster than a true one that lacks a recent precedent to hide behind.
Then a real announcement — with real caveats. Three days later, on September 8, 2026, OpenAI — not Anthropic — announced an actual claimed result: roughly 10,000 coordinating agents, running an unreleased internal model over 88 hours, produced an analytical proof and Lean formalization showing a smooth 3D fluid can develop a singularity in finite time. That's a genuine, detailed piece of work — 165 pages, a formal Lean 4 theorem statement, and disclosed process documentation nobody else in this space has matched.
But read the fine print, which OpenAI itself discloses: the result resolves the Clay Institute's forced-case variants (C and D) — where a smooth external force is injected to twist the fluid into a singularity — not the natural, unforced case most people mean when they say "solved Navier-Stokes." OpenAI is explicit that it is not claiming the $1 million prize for this reason. Public critics, most visibly X commentator Mehdi (e/λ), called using the forced-case route to imply "solved Navier-Stokes" a "major conceptual scam" — a sharp framing, but one built on a distinction OpenAI's own announcement already disclosed, not a new discovery.
The story then got messier, not clearer. NYU mathematician Tristan Buckmaster publicly alleged that OpenAI's effort was triggered by rumors of his own private, unpublished research with Anthropic researcher Levent Alpöge, and that the "achieved from just the problem statement" framing understated how much the effort had been steered. OpenAI's Sebastien Bubeck later confirmed in his own words that the effort did start because of viral rumors about Anthropic — settling that one specific point — while disputing the rest of Buckmaster's account. None of that credit dispute changes the math, but it's a second, independent reason to treat "AI solved Navier-Stokes" as a much messier sentence than a screenshot implies.
The net result for this problem: a real, technically detailed, Lean-verified proof exists — of a narrower variant than the prize actually asks about, produced under contested circumstances, not yet peer-reviewed, and not submitted to the Clay Mathematics Institute. "Recent AI claim, not yet verified" is the correct one-line summary, and it's also a template for how to read every future claim like it.
5. The Hodge Conjecture — untouched
The question: for certain well-behaved geometric spaces, which shapes can be described using purely algebraic equations, as opposed to more general topological or analytic methods? It's a bridge question between algebraic geometry and topology, and a "yes, here's the classification" answer would unify large parts of modern mathematics.
No AI system has publicly attempted this problem. It sits deep in algebraic geometry, a field where progress typically requires new conceptual frameworks rather than extending known numerical bounds — the same shape of difficulty that has kept Yang-Mills and Birch-Swinnerton-Dyer untouched. Nothing about the pattern in AI's actual wins this year (recombining known techniques on Riemann, formalizing an existing proof for Fermat) suggests this is close to changing.
6. Birch and Swinnerton-Dyer Conjecture — untouched
The question: for elliptic curves — equations of the form y² = x³ + ax + b — how many rational-number solutions exist? The conjecture predicts the answer is governed by the behavior of an associated L-function at a specific point. It's directly relevant to modern cryptography, much of which relies on elliptic curve arithmetic.
Like Hodge and Yang-Mills, there is no publicly reported AI progress on Birch and Swinnerton-Dyer as of September 2026. It's worth noting the same infrastructure that produced the Riemann bound — L-function analysis, Lean formalization — is at least in the same mathematical neighborhood as this conjecture, which is one reason it's plausible AI labs eventually attempt it. Plausible is not the same as attempted; nobody has announced work here yet.
7. Poincaré Conjecture — solved, by a human, in 2002
The question: is every simply-connected, closed 3-dimensional shape topologically equivalent to a sphere? Put more plainly: are there any "exotic" 3D shapes without holes that aren't secretly just spheres, viewed the right way?
This is the one Millennium Prize Problem that's actually been solved — and it's the sharpest possible contrast to the Navier-Stokes situation above. Grigori Perelman proved it across a series of papers in 2002-2003, extending Richard Hamilton's Ricci flow program with a technique called "surgery" to handle singularities that develop during the flow. It took the mathematical community years of careful, independent verification by multiple groups before the proof was widely accepted. The Clay Mathematics Institute awarded Perelman the $1 million prize in 2010. He declined it, and also declined a Fields Medal for the same work in 2006.
Notice what a real, accepted Millennium Prize solution actually looks like: years of expert scrutiny, multiple independent verification groups, a slow-building consensus — not a same-week announcement, not a viral thread, not a 165-page PDF with a press release attached. That's not a knock on doing work quickly; it's a reminder of what the finish line for "solved" actually requires, which is precisely the standard the September 2026 Navier-Stokes claim hasn't come close to meeting yet.
What "unsolved" and "unverified" actually mean here
Two different bars are getting collapsed together in casual reporting on all of this, and it's worth separating them explicitly:
- A Lean-checkable proof confirms a stated theorem follows logically from its assumptions. It's real, valuable evidence — stronger than a prose argument a human hasn't fully audited. It does not confirm the stated theorem is the same as the historical open problem, or that the assumptions weren't quietly weakened to make the proof tractable. That's exactly the gap in OpenAI's forced-vs-natural Navier-Stokes result: the Lean proof is real, but it certifies a narrower claim than the headline implies.
- Clay Mathematics Institute certification requires publication in a recognized journal, a two-year waiting period, and independent expert review before the $1 million prize is awarded — the process Perelman's Poincaré proof went through. No AI-generated Millennium Prize claim, on any of the seven problems, has entered that process as of this post.
If a future claim on any of these seven problems shows up in your feed, the questions worth asking before sharing it are the same three that separated signal from noise in the Navier-Stokes story above: Who is making the claim, using a named and disclosed method? Has anyone outside that organization examined it? Is there a machine-checkable artifact, and does it actually match the historical problem statement — not a redefined, easier version of it?
What this actually changes if you build with AI
The pattern across every genuine AI math result this year — the Riemann bound, the Fermat Lean formalization, even the disputed parts of OpenAI's Navier-Stokes effort — is the same one explainx.ai's broader look at AI and mathematicians argues for the field generally: AI is currently strongest at searching within a known space of techniques for a combination nobody had assembled, not at inventing genuinely new frameworks from a blank page. That's exactly why P vs NP, Hodge, and Yang-Mills — problems with no known "almost there" partial result to extend — remain completely untouched, while Riemann, which had 112 years of incremental human bounds to build on, got a real jump.
The practical takeaway for anyone running agent systems on hard problems isn't "wait for AGI to solve P vs NP." It's narrower and more useful: agent swarms are currently best deployed against problems that already have a partial solution path and a way to mechanically check the output — a Lean formalization, a numerical counterexample search, a test suite. When a problem has neither, as with four of these seven, throwing more agents at it hasn't moved anything yet, and there's no evidence in this year's results that it will.
What people are asking
Is any Millennium Prize Problem close to being solved by AI? The Riemann Hypothesis has the most real AI-touched progress (the 67.2% bound), but Anthropic itself says that specific technique won't lead to a full proof. Navier-Stokes has the most publicized claim, but it resolves a narrower forced-case variant than the actual prize question. Neither is "close" in the sense of being one more push away.
Why did the Navier-Stokes rumor about Claude turn out to be wrong? It traced back to a single X user's explicitly stated prediction, not a real Anthropic announcement — and per NYU mathematician Tristan Buckmaster's later public statement, it most likely originated from secondhand, misattributed knowledge of his own real (but non-Millennium) private research, not any actual Anthropic project.
Does OpenAI's Navier-Stokes result count as solving the Millennium Prize Problem? No, by OpenAI's own account. It solves the forced-case variant using an externally applied force, and OpenAI explicitly states it is not claiming the Clay Institute's $1 million prize for this result.
Why hasn't AI touched Hodge, Yang-Mills, or Birch-Swinnerton-Dyer at all? Every real AI math win this year worked by extending a chain of existing human results with a clear numerical or symbolic target. These three problems don't offer that kind of incremental path — they require new conceptual frameworks, which is where current AI systems are weakest.
Who decides if a Millennium Prize Problem is officially solved? The Clay Mathematics Institute, which selected the seven problems in 2000 and requires publication, a two-year wait, and independent expert review before certifying a solution and awarding the $1 million prize — the process Grigori Perelman's Poincaré proof completed in 2010.
Related reading on explainx.ai
- Did Claude Solve Navier-Stokes? The Millennium Prize Rumor, Fact-Checked — the unconfirmed rumor that started all this
- OpenAI's Navier-Stokes Proof Is Now a Credit and Data Dispute — the real, dated claim, its forced-case caveat, and the Buckmaster dispute
- Claude Pushed a Riemann Zeta Bound From 41.6% to 67.2% — Using 60 Subagents — AI's one genuinely verified Millennium-adjacent result
- Claude Wrote the First Machine-Checked Proof of Fermat's Last Theorem — a full formalization of an existing human proof, not a new discovery
- Will AI Replace Mathematicians? The IEEE "Big Mathematics" Debate — the broader trajectory this fits into
- OpenAI's Models Resolved an Erdős Planar Unit-Distance Problem — a genuinely solved, non-Millennium open problem, for contrast
- RSA-260 Factored — Verify It Yourself — applied cryptographic progress adjacent to P vs NP
- Fact-Checking the Zhi-Wei Sun / GPT-5.6 Sol Prime-Gap Claim — another viral AI-math claim that needed the same scrutiny
Sources: Clay Mathematics Institute, "Millennium Prize Problems," claymath.org · Anthropic, "Learning more about Claude's mathematical capabilities," August 10, 2026 · OpenAI, "On the Navier-Stokes Millennium Prize Problem" and "Finite Time Blowup for Navier–Stokes" (PDF), September 8, 2026 · Tristan Buckmaster, public statement, cims.nyu.edu, September 2026
This post reflects AI math progress and claims as publicly known as of September 10, 2026. None of the six currently unsolved Millennium Prize Problems has been certified solved by the Clay Mathematics Institute. It will be updated if that changes.