Anthropic strikes back: Alpöge claims to have proven the Riemann hypothesis with Claude, Buckmaster publishes a Navier-Stokes blowup without external force
🔎 Two mathematical bombs, one lab, twenty-four hours
OpenAI had taken Navier-Stokes. Anthropic has just responded — twice, on the same day. This Friday, November 20, 2026, Levent Alpöge, a mathematician recruited by Anthropic, publishes a 4-page note claiming that Claude helped him prove the generalized Riemann hypothesis, with a Lean 4 formalization presented as machine-verifiable. A few hours later, Tristan Buckmaster (NYU) and Alpöge submit a 79-page paper claiming a finite-time blowup of the 3D Navier-Stokes equations — without external force — backed by 600,000 lines of Lean 4 (AI Weekly, Scientific American, AI Weekly).
Two of the hardest problems in all of mathematics. A single lab. The same day. After OpenAI's 88 hours and 10,000 agents on Navier-Stokes, the war between AI math labs has just changed scale.
One piece of advice above all: keep your cool. No independent certificate, no peer review at this stage. These are claims, not theorems. This article sorts out what is claimed, what is verifiable, and what belongs to wartime PR.
The Essentials
- Claim #1 — Generalized Riemann: a 4-page note by Alpöge (Anthropic) claims that Claude helped prove the generalized Riemann hypothesis, formalized in Lean 4. No independent verification to date.
- Claim #2 — Navier-Stokes: Buckmaster (NYU) and Alpöge claim a finite-time blowup of the 3D equations without external forcing, in a 79-page paper accompanied by 600,000 lines of Lean 4. Also unverified.
- Status: no certificate, no peer review. We treat both as claims, period.
- Stakes: the blowup, if confirmed, checks one of the two outcomes of the Millennium Problem ($1M); GRH would restructure part of algorithmic number theory.
- Context: a direct response to OpenAI, which had announced Navier-Stokes in 88 hours with 10,000 agents and a public Lean proof, then targeted the Hodge conjecture.
Recommended Tools
| Tool | Main use | Price (November 2026) | Ideal for |
|---|---|---|---|
| Claude | Exploration, writing, assisted formalization | Pro $20/month, Max up to $200/month (check on anthropic.com) | Researchers, mathematicians |
| Lean 4 | Proof assistant, machine verification | Free, open source | Formalizing and auditing proofs |
| mathlib | Lean's standard mathematical library | Free | Reusing over a million lines of verified math |
| Zulip Lean | Community tracking of verifications | Free | Auditing claims in real time |
What Exactly Anthropic Claims
Direct answer: two distinct documents, published on the same day, claiming two major results with AI as the driving force — and, for now, no third-party-verifiable proof.
The first document is a 4-page note signed by Levent Alpöge. It claims that the generalized Riemann hypothesis has been proven, with the help of Claude, and that the proof has been formalized in Lean 4. Scientific American calls Riemann "the most famous problem in all of mathematics" — and the generalized version is an even stronger form of it, hence a heavier claim.
The second is a 79-page paper co-signed with Tristan Buckmaster, a fluid dynamics specialist at NYU. It claims a finite-time blowup of the 3D Navier-Stokes equations without external forcing, backed by 600,000 lines of Lean 4 (AI Weekly).
A detail that matters: Alpöge signs both. On the Anthropic side for Riemann, on the academic side for Navier-Stokes. He is the bridge between the lab and the university — and the central figure of this counterattack.
Two claims, one table
| Element | GRH note (Alpöge) | NS paper (Buckmaster–Alpöge) |
|---|---|---|
| Document length | 4 pages | 79 pages |
| Claimed role of AI | Claude "helped" | AI-assisted |
| Lean 4 formalization | Announced | 600,000 lines |
| Independent certificate | No | No |
| Peer review | No | No |
| Stakes if true | GRH proven | Clay Prize ($1M) |
The claim/proof distinction is the only lens that matters here. A claim can be brilliant or empty; only the machine — or a community of experts — can settle it. As I write these lines: neither one nor the other.
Generalized Riemann: a 4-page claim that carries the weight of 167 years of failures
Direct answer: if true, this is the greatest result in the history of AI-assisted mathematics — because 167 years of attacks have only ever produced a fraction of the zeros on the critical line.
A useful reminder. The Riemann hypothesis (1859) states that all non-trivial zeros of the zeta function have real part 1/2. The generalized version (GRH) extends the statement to Dirichlet L-functions — and it is this version that carries most of the algorithmic consequences.
The state of the art, meanwhile, is disarmingly modest: one third of the zeros proven on the line (Levinson, 1974), 40% (Conrey, 1989), about 41.7% at the best published rate (2020). No one has ever come close to 100%. A wall of partial results — and then, out of nowhere, a 4-page note.
The format is surprising, but it has an internal logic: the short note lays out the strategy, the machine verifies the details. This is the model of Scholze's Liquid Tensor Experiment, whose key theorem was certified in Lean between 2021 and 2022. Except there's one major catch: mathlib, Lean's standard library (over a million lines in 2023), poorly covers advanced analytic number theory — zero-free regions, explicit formulas, density bounds. Formalizing a complete GRH would require months of unprecedented formalization. The skeptics' question #1 therefore boils down to one sentence: where is the repository, and what does it actually contain?
What "Claude helped" means — and doesn't mean
The note remains vague about the model's exact role. Three scenarios, three radically different meanings: Claude as an exploration and brainstorming tool; Claude as a generator of proof candidates that the human sorts through; Claude as the effective co-author of the key steps, all the way up to formalization.
No public details for now. And that's where the rub lies: "helped" is the most flexible word in AI's marketing vocabulary. The truth will be in the Git repository, not in the note.
Navier-Stokes without external force: the honest version of the Clay prize
Direct answer: this is exactly the formulation of the Millennium Problem — and so, if the proof holds, a Million Dollar Problem falls, via the counterexample side.
The problem, posed by the Clay Mathematics Institute in 2000: do smooth solutions of the 3D Navier-Stokes equations, without external force, remain smooth for all time, or can they blow up in finite time? Prove regularity or construct a blowup: $1,000,000 on the line. Of the seven Millennium Problems, only one has been solved — Poincaré, by Perelman (2002-2003).
What already existed, and this matters: global weak solutions whose uniqueness remains open (Leray, 1934); non-uniqueness of weak solutions via convex integration (Buckmaster–Vicol, Annals of Mathematics, 2019); a blowup for an averaged version of the equations, hence a cousin problem (Tao, 2014, detailed on his blog); and blowups for Euler at low regularity (Elgindi, 2021). Nothing, ever, for smooth 3D Navier-Stokes, without forcing.
Hence the importance of the phrase "without external force". With forcing, you can inject energy and manufacture blowups in nearby settings. Without forcing, you attack the physical equation itself — the one that actually describes fluids. That's the honest version of the problem, and the hardest one. It's also what sets this claim apart from previous announcements on the subject.
The 600,000 lines of Lean 4 give a measure of the scale: mathlib as a whole passed one million lines in 2023, after a decade of community work. A PDE paper formalized end to end would, on its own, be one of the largest formalization projects ever completed. Irony of the timeline: if the file is clean, the mechanical verification itself will take hours — not months.
Lean 4: the machine referee, its guarantees and its blind spots
Direct answer: Lean turns "do you believe this mathematician?" into "does this file compile?" — but only if the theorem statement is the right one, and that's where everything hinges.
The principle: Lean 4 is a proof assistant in which every step must be justified all the way down to the basic axioms. A file that compiles proves its statement, period. A Mochizuki-style "gap" is structurally impossible in a machine artifact: the gap, if it exists, prevents compilation. That's the whole difference from a proof on paper.
But three blind spots remain, and they are precisely where a claim can dissolve:
- The
sorry: in Lean, a hole in a proof is markedsorry. A file riddled withsorrycompiles, but proves nothing. Every auditor's first reflex: search for the word. - The axioms:
#print axioms theorem_namelists the dependencies. Lean rests on three standard axioms (propext, Classical.choice, Quot.sound). Any axiom added "by hand" — for instance, a statement equivalent to GRH admitted as a hypothesis — ruins the claim. - The statement itself: the favorite trap. A blowup for rough data is not a blowup for smooth data. GRH for a subfamily of L-functions is not GRH. A formal theorem can be formally true and mathematically disappointing.
That's why experts will read the statement first, not the proof. Half a page of theorem is worth a thousand lines of mechanical verification.
And where does AI fit into all this? Formalization is, literally, code. The models that dominate code — Claude Opus 4.7 (Adaptive), GPT-5.5, Gemini 3.1 Pro — are structurally the same ones capable of writing Lean at scale. If you want a sense of their respective strengths, our guide to the best LLMs for coding details what each one can do. And let's not forget that formal AI didn't come out of nowhere: AlphaProof took silver at IMO 2024, and the project to formalize Fermat's Last Theorem was launched in 2025.
The verification checklist, point by point
- Does the repository compile? Pinned toolchain, reproducible build. Without a public repository, everything else is hot air.
- Zero
sorry, standard axioms?#print axiomson the final theorem, not on a carefully chosen intermediate lemma. - Does the statement match the claim? Full GRH? Blowup for smooth data, with no external forcing, in finite time?
- Who re-verifies? An independent re-formalization, even a partial one, remains the gold standard.
OpenAI vs Anthropic: the war of the AI math labs
Direct answer: these two announcements are not a calendar coincidence, it's a timed counterattack.
The timeline speaks for itself. OpenAI announces Navier-Stokes solved in 88 hours with 10,000 agents and a public Lean proof — which we dissected here. OpenAI follows up by taking aim at the Hodge conjecture (our analysis). And Anthropic responds with two claims on the same day, signed by an in-house mathematician and a renowned fluid dynamics specialist.
Anthropic doesn't come empty-handed to this match: its agent Claude Fable 5.2 had already solved two open black hole problems — the Hawking page (2003) and an exact modular bootstrap ceiling (our article).
Two philosophies clash, and they are readable in the numbers. OpenAI: brute force, 10,000 agents in parallel, 88 hours. Anthropic: a leading mathematician plus a cutting-edge model, a 79-page paper. My opinion, without beating around the bush: for a theorem of this depth, the second approach is more credible — new ideas don't parallelize. The first is more spectacular. Both, for now, are worth exactly the same price on the proof market: zero, as long as nothing is verified.
Why are the labs fixated on math? Three reasons, all strategic:
- It's the only unbreakable benchmark. A code score can be inflated through training data contamination; a Millennium theorem cannot. Machine verification closes the door to rigged benchmarks.
- It's a talent magnet. The best mathematicians want the best machines. Alpöge at Anthropic is a signal addressed to an entire generation of researchers.
- It's very high-intensity marketing. "Our AI helped prove Riemann" is worth all the leaderboards in the world — if it's true.
The collateral consequence is academia. The Lean community counts only a few hundred active contributors; auditing 600,000 lines and a GRH statement will tie up the best of them for weeks. Universities have neither the computing resources nor the staffing to keep up with the pace of the labs. Verification — the link that gives everything else its value — is running behind. And while the labs face off over Riemann, the question that interests most readers remains more prosaic: which model should you choose for your own work? Our Claude vs ChatGPT comparison remains the best starting point.
Riemann = broken crypto? No. What it would actually change
Straight answer: no, RSA isn't falling, your private key won't budge by a single byte — and that's the most stubborn myth behind every wave of Riemann announcements.
The GRH factorizes nothing. RSA's security rests on the difficulty of factoring, not on the possibility that Riemann might be false. What the GRH would change is the algorithmics of numbers: Miller's deterministic primality test (1976), conditional on the GRH, would move from "heuristic" to "proven" status — even though AKS (2002) already provides an unconditional deterministic test; Bach's explicit bounds (1990) would make rigorous a whole swath of algorithms on groups and number fields. Real gains, but in constants and guarantees — not a cryptographic revolution.
On the fluids side, a verified blowup won't change anything about your CFD simulator tomorrow morning. The implications — turbulence, singularities, the foundations of PDEs — will be measured in decades. Above all, it would be an intellectual earthquake: the first resolution — a negative one — of a Millennium Problem, built with AI.
The real product of these announcements lies elsewhere: the workflow. A leading mathematician, a language model, a proof assistant, a verifiable artifact. If this pipeline works, it will generalize to entire swathes of research — and paradoxically, the big winners of this war may not be the labs, but the verification tools.
❌ Common mistakes
Three mental shortcuts are already showing up in every discussion — all three are wrong.
Mistake 1: taking the press release for the proof
What's wrong: the history of mathematics is a graveyard of claims never converted. Mochizuki's proof of the abc conjecture remained contested for a decade before a divisive publication; Atiyah's Riemann "proof" (2018) evaporated within weeks. The solution: wait for the artifact. No audited Lean repository = no theorem.
Mistake 2: believing that 600,000 lines of Lean guarantee the result
What's wrong: the Lean lines prove the statement as written, not the one you think you're reading. A weakened statement or an ad hoc axiom is enough to hollow out the claim. The solution: check the theorem statement and the axioms before marveling at the volume of code.
Mistake 3: "Riemann solved, so cryptography is dead"
What's wrong: confusing the Riemann hypothesis with factorization. Nothing in the GRH breaks RSA or any deployed encryption. The solution: follow the real impacts — algorithmic number theory, primality tests, explicit bounds — and ignore the alarmist headlines.
❓ Frequently Asked Questions
Has Claude proven the Riemann Hypothesis?
No. Alpöge claims that Claude helped him construct a proof of GRH, formalized in Lean 4. As long as no repository has been independently audited — with no sorry and no non-standard axiom — it remains a claim. Recent history (Atiyah in 2018) calls for maximum caution, even when the author is a serious mathematician.
What exactly is the generalized Riemann Hypothesis?
The Riemann Hypothesis (1859) states that all non-trivial zeros of the zeta function have a real part equal to 1/2. The generalized version extends the statement to Dirichlet L-functions. It's this version that provides effective bounds in number theory — hence its algorithmic importance, well beyond the prestige of the name.
Does the Navier-Stokes blowup qualify for the Clay Prize?
Yes, if the proof is verified. A finite-time blowup for smooth data, with no external forcing, constitutes the counterexample anticipated by the problem statement. Mind the rules: the Clay Mathematics Institute requires publication in a peer-reviewed journal and roughly two years of community acceptance before any $1M check.
Which Claude model was used?
The announcement doesn't specify. Anthropic's current lineup runs from Claude Sonnet 4.6 to Claude Opus 4.7 (Adaptive), its most capable model for agentic reasoning. Without public confirmation, any precise attribution to a given model is speculation — including when it comes from a journalist.
How long does it take to verify 600,000 lines of Lean?
If the repository is public with a pinned toolchain: recompilation takes hours. Auditing the statement: a few days for specialists. Full community digestion: weeks to months — the Liquid Tensor Experiment took about 18 months for the entire project, with the key lemma certified as early as 2021.
Hasn't OpenAI already done better on Navier-Stokes?
The two announcements aren't directly comparable: 88 hours and 10,000 agents on one side, 79 pages and 600,000 lines of Lean on the other, with possibly different problem scopes. As long as neither proof is verified, the only honest comparison fits in five words: two claims, zero certificates.
✅ Conclusion
Two historic claims, zero verified proof: the ball is now in Lean's court, in the mathematical community's — and in the clocks of OpenAI and Anthropic, which have just made verification as competitive as the announcement. To follow what comes next, start with our breakdowns of the OpenAI announcements, and keep one rule in mind: in math as in code, as long as it hasn't compiled under a third party's eyes, it isn't proven.