88 Hours, 10,000 Agents and a Millennium Problem
The claim, the Lean proof, and the credit war behind it

What OpenAI actually claims

An internal system produced an analytical proof and a Lean formalization showing that an initially smooth fluid at rest, subjected to a smooth external force, can develop a singularity in finite time while its total energy stays bounded throughout. OpenAI says this establishes statement "C" (and also "D") in Charles Fefferman's official Clay formulation — i.e. a disproof of global regularity, rather than a proof that solutions always stay smooth. It explicitly says it will not claim the $1 million prize.

The one discipline this report enforces: three different statuses

Our verdict

• Mathematical importance, if it holds: landmark — the most significant result yet to involve AI, and only the second Millennium Problem ever answered (after the Poincaré Conjecture).
• Current certainty: medium, leaning "to be verified." Machine-checking raises the floor substantially, but statement equivalence, the relationship between the two teams' routes, and attribution are still being sorted out.
• Paradigm signal: strong. Whether this exact proof survives or who signs it, "massively parallel agent search + formalized verification" has now been shown to act on an open frontier problem — not only on problems with known answers.
• Near-term productizability: low. This was a multi-million-dollar, capital-intensive run on a model nobody outside can access — a scientific instrument, not an API capability you can buy by the call.

• Not "OpenAI first, others chasing": the other side's key Lean-verified breakthroughs date to mid/late August; OpenAI itself dates its start to the Sep 1 rumor. Two teams hit a very narrow niche route almost simultaneously.
• They did not prove the identical proposition even in the warm-up: in Euler, Buckmaster–Alpöge proved the forced (smooth forcing) version; OpenAI's agents proved the unforced version. At the Navier–Stokes level only OpenAI claims a complete result; the other side reached an unverified hypo-dissipative NS result and chose not to release it.
• In its "Concurrent work" section OpenAI cedes priority on 3D Euler — "we recognize the priority of their work on forced Euler and congratulate them" — while claiming Navier–Stokes for itself.

3.1 The question in one paragraph

The Navier–Stokes equations apply Newton's second law (F=ma) to a fluid treated as a continuous medium, rather than as a swarm of molecules; they underwrite aircraft design, weather forecasting and the study of blood flow. Written down by Navier (1822) and Stokes (1845), one basic question stayed open: given a three-dimensional, initially perfectly smooth incompressible fluid, do solutions stay well-behaved forever, or can some infinitesimal parcel accelerate without bound in finite time — a singularity? Viscosity normally damps motion, which is what makes the question hard: the equations would have to blow up through the fluid's own dynamics, not because an infinite force was inserted by hand. OpenAI's answer is that they can: finite-time blowup.

3.2 Why it took ninety years

The final hurdle, put most clearly by Quanta: Córdoba and Martínez-Zoroa build an infinite sequence of non-singular "layers" and combine them in an infinite cascade that contains a singularity. Each layer is smooth, but stacking them used to leave the forcing function with undesirable properties — failing the Clay requirement that forcing be smooth. The step both AI teams claim to have crossed is precisely producing a cascade that yields a singularity while still leaving a smooth force behind.

3.3 What the blowup looks like — and what it does and doesn't mean for engineering

The constructed object is a vortex: it spirals inward and is axially stretched like spaghetti; its core thins and spins faster and faster, going unbounded while total energy stays finite. The technical feat is that acceleration, pressure gradients, momentum transfer and viscosity must all grow large yet cancel with exquisite precision, leaving only a smooth external force.

4.1 Key figures (official figures vs outside estimates kept apart)

4.2 What is genuinely new in the method

This was not "type the problem into a model and wait." Different agent groups were handed different statement variants — A/B (which would yield a proof of regularity) and C/D (which would yield a blowup) — as an explicit hedge against betting on the wrong direction. Agents could read a cached copy of the internet and run code, and communicate within groups; Codex then "cross-pollinated" groups by consolidating the most useful intermediate results, feeding the earlier Euler solution back in as a prompt. The human role was to design the search space, decompose the problem, orchestrate the swarm and judge which path was promising, rather than perform the derivation. OpenAI goes so far as to say its team collectively lacked research-level expertise in fluid dynamics and "could not make a genuinely meaningful contribution to the mathematics" — a self-describing account that should itself be read with some caution.

4.3 What Lean guarantees — and the one check it cannot do for you

Lean is an interactive proof assistant: propositions and proofs are written in a formal language and the software checks every inference. A proof that compiles, uses no sorry placeholder and rests only on Mathlib plus standard axioms has a closed logical chain within the definitions and axioms it chose. That is a full notch more trustworthy than a handwritten preprint awaiting review — and the essential difference from the graveyard of failed NS proofs.

• ① Statement correctness is not automatic: whether the formal encodings of "smooth forcing," "finite energy," "all of ℝ³ / the torus" exactly match Fefferman's C/D conditions must be checked line by line by outside experts. OpenAI claims to cover C and D on both ℝ³ and the 3-torus; that claim is precisely what needs auditing.
• ② It certifies neither physical meaning nor significance: the machine is silent on what the result means; humans supply that.
• ③ It is not yet community acceptance: outside mathematicians have not independently built and audited the repo. Formalization makes the result auditable; auditable ≠ accepted.

5.1 What the other team did

NYU professor Tristan Buckmaster and mathematician Levent Alpöge — employed by Anthropic but at pains to say he participated personally, with no institutional project and tools paid from Buckmaster's own research funds — used Anthropic Claude, OpenAI Codex, GPT-5.6 Sol and, later, Astra (Astra only for writeups and auditing) to obtain Lean-verified finite-time blowup with smooth forcing for three model equations: the incompressible porous medium equation (IPM, also with Matei P. Coiculescu), the 2D Boussinesq system, and 3D incompressible Euler. They also believe they have blowup for hypo-dissipative Navier–Stokes but declined to release it because the Lean verification was unfinished and no presentable writeup existed.

Fields medalist Terence Tao, writing Sep 7, called it an "exciting" / "remarkable" advance along the Córdoba–Martínez-Zoroa route, pushing subcritical/rough forcing all the way to fully smooth, and judged extension to full Navier–Stokes "very feasible… in the near future" with "no fundamental obstacle." He also added a cooler note worth keeping: solving is only a proxy goal; the primary value is mathematical understanding, and without it, the NS regularity problem is of "far less intrinsic significance… than is sometimes promoted in popular media."

5.2 The two teams at a glance

5.3 Three layers of the quarrel, ordered by evidence strength

Layer 1 — "Did OpenAI sprint after a leak?" (timeline largely checkable; interpretations differ). Buckmaster says the pair were forced to publish early once their progress leaked; OpenAI admits it began Sep 1 only after the "Anthropic solved it" rumor and reached out after completing its work Sep 6 to propose a concurrent release and let the other side publish first. The timeline itself is largely verifiable — the other side's key results are mid/late August, OpenAI's run is Sep 1–6.

Layer 2 — Were their private Codex drafts used in training? (the crux; no evidence either way; OpenAI did not fully answer). The pair kept every draft, prompt and intermediate in Codex. Buckmaster's question was never "did the model read my session live" but "did those sessions enter the training/improvement pipeline." He says he was told "the model does not look up user data," while his follow-up on training "did not get an answer." OpenAI's own wording: no specific user data was accessed to solve the problem, but it "cannot rule out that de-identified data derived from their usage of our products helped improve our models," though it considers that unlikely.

Layer 3 — Authorship and alleged pressure (he-said/he-said; each side has statements or screenshots). Buckmaster says two proposals were put to him, one of which would have him sole-author an NS paper that omitted Alpöge because of his Anthropic employment, and recounts language such as "Why would you ruin your career?" Bubeck publicly denies ever asking to remove Alpöge, concedes saying "if Levent were not an Anthropic employee, things would be simpler" on IP grounds, apologizes for "terrible wording" he says he retracted on the spot, and counters that the call opened with defamatory accusations and a threat to contact the press. Sam Altman likewise says he had wanted a joint release, offered to let the other side publish first and pursue the prize, and could not extend the same arrangement to an Anthropic employee. This layer currently consists of conflicting single-party accounts; the honest move is to set them alongside each other.

5.4 The overlooked "real heroes"

What the quarrel actually clarifies is that both AI teams stand on the same two human mathematicians: Diego Córdoba (ICMAT, Madrid) and Luis Martínez-Zoroa (CUNEF). Fefferman, who wrote the official Clay problem, says flatly "the heroes of the story are Córdoba and Martínez-Zoroa"; Buckmaster writes that Martínez-Zoroa deserves a Fields Medal for the body of work; Córdoba jokes, "I don't use AI: I have Luis." The reminder is blunt: AI rapidly pushed a niche route that humans had opened across its last technical gap — it did not invent the research program.

6.1 The Clay rules

A candidate must satisfy all three: (i) publication in a "qualifying outlet" — a peer-reviewed journal of world-wide repute; (ii) at least two years of standing after publication; (iii) general acceptance by the global mathematics community. Direct submissions are not accepted. For scale: the only Millennium Problem ever solved, the Poincaré Conjecture, ran from Grigori Perelman's November 2002 preprint to the March 2010 award — over seven years — and he declined the prize. As of writing, the Clay site still lists Navier–Stokes as unsolved, the institute has neither accepted nor rejected OpenAI's claim, and OpenAI says it will not apply for the $1 million, framing the release as "a snapshot in time… not a culmination."

6.2 Why the profession's slow, procedural caution is correct

Navier–Stokes is a famous graveyard of proofs: more than one announced solution later collapsed under scrutiny — Otelbaev's 2014 claim, tracked publicly by Tao among others, was ultimately judged unsalvageable. Facing a 165-page proof from a non-traditional channel wrapped in commercial and priority disputes, the discipline's instinct to independently build the Lean, read slowly and withhold judgment is not conservatism; it is the mechanism that keeps mathematics reliable.

• Leap one (2024): medal-level performance on olympiad problems that have standard answers — tests problem-solving.
• Leap two (2025 – first half 2026): start on known open problems (the Erdős legacy) — tests advancing inside terrain humans had already mapped.
• Leap three (Sep 2026): face an open frontier with no ready path, where you must even bet between opposite directions (prove-smooth vs construct-blowup), search at massive parallelism, then verify formally — tests helping open up new terrain.

8.1 The research production function changes: scarcity moves from solving to asking and orchestrating

When ten thousand agents can try hundreds of proof paths in parallel, the scarce resource shifts away from grinding through long derivations and toward the other end: reformulating a fuzzy scientific question into something a machine can search and verify (Tao's "good questions become the scarce resource"); designing the search space (A/B/C/D variants, easier warm-ups, cross-group pollination); using formalization as a backstop while humans perform the hardest step of confirming formalized-problem-equals-real-problem; and rewriting machine "AI slop" into arguments a community can read and inherit — Buckmaster's apology for rough drafts is evidence this step cannot be skipped. High-value research roles look increasingly like research product manager + conductor of an agent swarm + formalization auditor, rather than the lone derivor.

8.2 A new arena for frontier competition: not chat scores but "who is relied on at the real frontier"

• The two routes leaned heavily on Anthropic Claude and OpenAI Codex / GPT-5.6 Sol / GPT-6 Astra, plus OpenAI's unreleased next model. On the hardest agentic long-horizon tasks — multi-file, long chains, continuous self-verification — which tools researchers actually vote for with their feet says more than any chat benchmark.
• The decisive capability axes this episode re-names: long-horizon agent stability, large multi-agent orchestration, coordination with formalization/code tools, verifiability of output, and the speed of model generations (OpenAI pointedly flags an internal model "significantly stronger than GPT-6 Astra and still training" — both a capability preview and a commercial signal).
• For model selection: alongside static scores, evaluate long-horizon task completion, tool/verification-chain coordination, and controllable, auditable output. Stay clear-eyed that top research capability sits a generation and a release-lag away from the commercial models an enterprise can buy.

8.3 A capital-intensive scientific instrument — not yet a cheap API capability

Several million dollars, ~10,000 concurrent agents, ~130 billion output tokens, 88 hours, and an unreleased model you cannot buy: this is closer to running an experiment on a particle collider than calling an API. It will not become a standard, affordable service for SMEs any time soon. What productizes first — and far more certainly — is the decomposed parts of the pipeline: formal-verification tooling, code/research agent orchestration frameworks, auditable research-collaboration environments, and vertical "multi-agent search + verification" workflows.

8.4 A rules vacuum: attribution, data and reproducibility norms lag behind capability

• Authorship and credit: how is credit divided among the model, the company that built it, the researchers who steered the agents, and the humans who opened the route? How is joint release even possible when a participant works for a rival?
• Data boundaries: whether — and under which opt-out settings — a vendor may improve models on unpublished work fed into its tools needs far clearer defaults and technical guarantees than exist today.
• Reproducibility: when a key result rests on an unreleased model, how does the community independently repeat it? Is "public artifact + black-box process" enough to constitute a discovery?
• Evaluation: when machine-verified proof becomes routine, peer review, journals, hiring and prizes all have to be redesigned.

None of these will be settled by this one episode, but it functioned as a stress test that handed the bill, in advance, to the entire academic and industrial system.

In the first week of September 2026, what genuinely crossed a threshold was not the result "AI solved a Millennium Problem" but the production method — massively parallel agent search, formalized verification as backstop, humans defining the problem — running a complete closed loop on a Millennium-class open question for the first time.

Toward the mathematical result itself we recommend the same patient agnosticism as the profession: Lean backing gives it a high floor, but until independent reproduction, statement-equivalence audit, peer review and the two-year test, the rigorous phrasing stays "a proof awaiting verification," and the Clay page still reads unsolved. The priority and data dispute should be decided by evidence; for now it is, at most, two accounts standing side by side.

Watchlist (falsifiable)

• Independent build: does an outside team successfully compile OpenAI's Lean repo and verify, line by line, that the formalized definitions are equivalent to Fefferman's official C/D?
• Expert reads: public judgments from in-field mathematicians (Córdoba, Martínez-Zoroa, Silvestre, Tao and others).
• Route convergence: do the pair's forced-Euler / hypo-dissipative NS and OpenAI's unforced-Euler / full NS ultimately prove to be the same path or materially different?
• Reproduction on public models: can the swarm attack be rerun on models anyone can access, rather than OpenAI's unreleased one? This decides how transferable the paradigm is.
• Clay's position: whether, and when, the "unsolved" label changes.
• New evidence on data/authorship: emails, an authoritative reading of Codex data terms, anything that turns the single-party accounts into checkable fact.

Three more certain takeaways for practitioners: the capability signal is real — "AI can only do exercises, not research" is obsolete; the human role changed but became more decisive — good questions, search-space design, equivalence checks and human-readable insight are the moat; and for go-to-market, decompose and temper expectations — the full run is a capital-intensive instrument, while verification, orchestration and audit are the parts that productize first, alongside judging models on long-horizon agents, tool coordination and verifiable output.