By Jake Hari. Kalshi prices and order-book sizes were fetched September 9, 2026, at 9:13 p.m. ET. Every claim about the proof itself is linked to the primary source.
On the evening of Tuesday, September 8, 2026, hours after OpenAI said an unreleased model had produced a proof for the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems, Kalshi opened a market beside it. The market does not pay on that proof. Read the contract and Navier-Stokes is the one open problem, alongside the long-solved Poincaré conjecture, that is written out of it by name. What traders are pricing is whether OpenAI announces a second solution. As of Wednesday night, buyers were offering 34 cents on the dollar and sellers were asking 39 for it to happen by New Year's, and the contract for the year after sat at a coin flip.
Most of the coverage this week is about whether the first one is real. For this contract, that question does not matter. What does matter is one sentence in OpenAI's own account of the 88 hours: the company says it pointed the same system at every open Millennium problem first, then pulled the agents off five of them to chase this one.
Stay ahead of the markets.
Daily insights and expert picks on Kalshi, Polymarket, and what's moving markets.
Free forever. Unsubscribe anytime.
The Quick Answer
The Kalshi market titled "Will OpenAI announce another Millennium Prize solution?" pays out only if OpenAI announces a complete solution to one of the five qualifying problems before its deadline: the Riemann Hypothesis, P versus NP, the Hodge Conjecture, Birch and Swinnerton-Dyer, or Yang-Mills. The Navier-Stokes result is excluded by name, so anyone buying because that proof looks solid has bought the wrong question. The Clay Mathematics Institute's verdict never enters into settlement either. Two facts most headlines skip: OpenAI says it does not intend to claim the $1 million prize, and the proof is still unverified, pending review by the wider mathematical community. The full contract wording, the resting size behind each price (it is small), and the one line in OpenAI's post that bears on the next attempt are below.
What The Contract Actually Pays On
Kalshi runs two contracts under one question, one closing at the end of 2026 and one at the end of 2027, and both settle on an announcement rather than on a result being correct. The clause that decides everything is a single sentence in the rules:
"Clay acceptance, peer review, and a prize award are not required."
The event is a press release. Kalshi names OpenAI itself as the settlement source, so the moment OpenAI publishes a post saying it has a complete solution to one of the five qualifying problems, and expressly states that OpenAI developed it or names the outside researchers who collaborated on it, the contract settles Yes. Whether the wider mathematical community accepts the proof, and whether the Clay Institute ever writes a check, is somebody else's market.
Four more doors that a casual reader might assume are open are closed in the same rulebook. A proof by an independent ChatGPT user does not count, even if OpenAI publicizes it. Giving a university team model access or research credits does not make OpenAI a collaborator. Partial results, numerical evidence, and a formalization of a proof someone else already found do not count as a complete solution. And the problem the whole news cycle is about, Navier-Stokes, is written out by name, alongside the long-settled Poincaré conjecture.
The habit of reading the settlement clause before the price is the same one that runs through Kalshi's OpenAI IPO contract, which pays on a regulatory trigger (an effective S-1, a priced deal, or an assigned ticker) rather than a first trade, and through the AGI board, which pays when OpenAI publishes the words rather than when the thing arrives. On this one the gap between headline and contract is wider than either, because the headline is about a proof the contract explicitly ignores. Same event, different rulebooks is the standing lesson of this whole category.
The Board
The bid is what a buyer will pay and the ask is what a seller wants, and on a thin market the gap between them, and the handful of contracts behind each, is the whole story, so the table shows what is resting at each price.
| Contract | Yes bid / ask | Resting at the best bid / best ask | Contracts traded since open / still open |
|---|---|---|---|
| Announces Another Solution Before Jan 1, 2027 | 34¢ / 39¢ | 1 / 307 | 3,501 / 2,655 |
| Announces Another Solution Before Jan 1, 2028 | 50¢ / 51¢ | 1,319 / 2,660 | 411 / 207 |
Order book as of September 9, 2026, 9:13 p.m. ET. Both contracts opened at 6:00 p.m. ET on September 8. "Contracts traded" is a count of contracts, not dollars; "still open" is open interest, the number of contracts currently held. Re-run every number below at the price you are actually offered.
Everything on this page is an estimate of probability, market or model estimates, not predictions of fact and not financial advice. Kalshi is a CFTC-regulated exchange for event contracts (18+; availability varies by state), and this page recommends no position on either side.
The near contract is the one to stare at. A single contract sits at the 34-cent bid. Below it the real buying interest starts: 40 contracts at 33, then 266 at 32 and 500 at 31. Above it, 307 contracts are offered at 39 and 200 at 40, and then nothing until 62. The quoted market is five cents wide, and the "price" people will repeat from the Kalshi app is whichever side the last trade hit. Further out, the book tells you what the crowd is really made of: about 650 contracts are bid between 8 and 14 cents, and 1,107 are offered at 70, resting orders from people hoping to get filled on a panic in either direction. Its first trade printed at 22 cents at 8:58 p.m. ET on Tuesday. At 9:04 p.m. ET on Wednesday, about 265 contracts were bought in a single second, at prices from 33 to 39. Any chart that shows this contract jumping from the low 30s to 39 is showing that second.
The far contract has the stranger history. Its first trade, at 6:32 p.m. ET on Tuesday, half an hour after listing, went off at 98 cents, on the proposition that OpenAI does this again within 16 months. By 9:45 a.m. ET on Wednesday a 201-contract sale at 50 had taken the price there in one sweep, and 1,319 contracts now rest on the bid at 50 against 2,660 on the ask at 51. It is a thinner market by trading, with 411 contracts changing hands across 14 trades, but it has far more contracts resting at the best prices, and its one-cent spread is the only thing about this board that looks like a settled opinion.
In payout terms, the 39-cent ask turns $100 into about $256 if OpenAI makes a qualifying announcement before New Year's, and into nothing if it does not, before Kalshi's fee touches the win. The 51-cent ask on the far contract turns $100 into about $196 on the same math, with the deadline a year later.
The Two Facts The Headlines Bury
Both of them are priced at zero in this contract, by design. Kalshi wrote the rules so that a disputed, unreviewed, unclaimed proof and a celebrated one settle the same way, on an announcement about a different problem, so the useful question about each fact is what it does to that event. In every case the answer is nothing, with one exception.
OpenAI is not claiming the money. The company's post says it plainly: "We do not intend to claim the Millennium Prize for this result." Part of that is the Clay Institute's own clock. Its rules require a solution to be published in a qualifying outlet, to sit for at least two years after publication, and to have "received general acceptance in the global mathematics community" before the Institute considers it. Even if OpenAI published tomorrow and every mathematician alive signed off, that two-year wait would run past the far contract's expiry on January 1, 2028. Martin Bridson, the Institute's president, told AFP, in a wire report carried by BNN Bloomberg:
"The process of evaluation is deliberately unhurried, and we shall ensure that it is absolutely rigorous."
As of this writing, the Institute's Navier-Stokes page still carries one word on it: Unsolved. The contract does not care. Clay acceptance is named in the rules as something that is not required.
No independent verification has been completed. What OpenAI published is a written proof plus a machine-checkable version in the Lean proof assistant, posted to a public GitHub repository that anyone with the tooling can build. A Lean check confirms that every step follows from the statement as it was formalized. The human work that remains is confirming that the formalized statement is the Clay problem, and that is what the coming weeks are for. On the narrow question of whether an applied force is allowed, the official problem statement written by Charles Fefferman is explicit: the breakdown alternatives, statements (C) and (D), ask for "a smooth, divergence-free vector field" and "a smooth f(x,t)" for which no solution exists, and OpenAI says its result establishes exactly those two statements.
The early reactions split along a line that has nothing to do with the mathematics, and nothing to do with settlement either. Fefferman told Quanta he was "thrilled that the problem was solved" and called Diego Córdoba and Luis Martínez-Zoroa "the heroes of the story." Terence Tao, in remarks reported by Fortune, worried that "the indiscriminate strip-mining of open problems for solutions may destroy the ecosystem from which the next generation of mathematical techniques, problems, and practitioners would have developed." Tao's objection is the most serious thing anyone has said about this result, and it cannot move a single share, because the contract asks whether OpenAI does it again, never whether it should.
The priority dispute is the exception. NYU's Tristan Buckmaster and Anthropic's Levent Alpöge had been using AI models on a closely related fluid problem, and Buckmaster told Fortune he had asked OpenAI "whether the model had been trained on, or had access to, our sessions in Codex, into which we had been putting all our drafts for the whole of this project." OpenAI's Sébastien Bubeck answered in the same piece: "We did not use their prompts or proofs to prompt our models or direct our agents." OpenAI's own post goes one step further than that denial, conceding that "while unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models," while noting the two Euler results differ: OpenAI's agents proved the unforced version, Buckmaster and Alpöge the forced one. Buckmaster also alleged that Bubeck pressed him personally over the dispute; Fortune has those quotes.
The dispute matters for the contract because of who gets named. The same two researchers, or any outside team, could announce a complete solution to one of the five qualifying problems, and this market would still settle No unless OpenAI's own announcement expressly identified them as collaborators on that solution. After this week, a joint announcement between the two camps is the last thing a trader should assume.
What the market is pricing, then, is repetition.
The Sentence In OpenAI's Post That Matters Most
OpenAI's account of how the proof happened contains the line that bears on this contract more than any mathematician's opinion does. The effort began on September 1, after a rumor that two Millennium problems had fallen, and the company writes:
"We launched an effort to evaluate it on all open Millennium Prize problems and a few other high-impact problems."
All of them. The five problems this contract pays on were on the same system's desk from day one, in parallel groups of agents, each fed "different variants of the problem statement." Then a smaller group of about 100 agents cracked a related problem, the unforced version of the same breakdown question for the Euler equations, in roughly 50 hours, and OpenAI made a choice: "we shifted agents away from the other Millennium Problems" and pointed them at Navier-Stokes, which the Euler result had made look most promising. The group that finished the job ran on the order of 10,000 concurrent agents and landed the result 88 hours after the first agents were launched on September 1. Along the way it sent 2.7 million messages and used about 130 billion output tokens, and OpenAI's GPT-6 Astra model then spent a further 17 hours formalizing the result in Lean.
You can read that history two ways, and the 34-to-39 spread on the near contract is roughly the argument between them. One reading: the other five problems had agents on them for days without a result the company chose to keep chasing, and OpenAI pulled resources off them for the one where a disproof, an explicit example of the equations breaking down, was within reach. Córdoba's line to Quanta, "Ten years ago, nobody believed there was a singularity for Navier-Stokes," is a reminder that the field had already turned toward breakdown; the agents found a path the community had started to suspect existed. Riemann, P versus NP, Hodge, Birch and Swinnerton-Dyer and Yang-Mills carry no such consensus about which way they break. The other reading: OpenAI now has a working template, a "further trained" model that improved during the run, and a very public reason to spend the "several million dollars" Bubeck put on the first run, in his remarks to Quanta, doing it again before the story cools.
One more wrinkle cuts against the second reading: partial results do not count. A dramatic step toward Riemann, of the kind that would dominate a news cycle, would not qualify. The contract needs a complete solution, so a headline-grabbing partial result leaves it exactly where it was.
How To Read A Book This Thin
The first trap is the spread. The near contract's quoted market is 34 bid, 39 ask, with one contract at the bid. Anyone who tells you "the market says 39%" is quoting the last price a buyer paid, and anyone who says "34%" is quoting a single resting contract. The honest statement is a range: somewhere in the mid-30s, with a five-cent toll for anyone who wants to buy and sell in the same afternoon. What a bid-ask spread actually costs is the arithmetic in full, and the short version is that a five-cent spread on a 35-cent contract is about a seventh of the position gone before anything happens.
The second trap is volume. About 3,500 contracts have traded on the near contract since it opened, and 411 on the far one. Thirty-five hundred contracts is a small crowd, and the opinion of a small crowd is a starting point, not a probability. The far contract shows what a small crowd's book looks like underneath: on each side, a few thousand contracts sit at 1 and 2 cents, lottery orders that will only fill in a collapse, and the sizable orders away from the inside are 1,413 contracts bid at 21 cents and about 3,000 offered at 90. Strip those out and the whole 16-month question rests almost entirely on the 1,319 bid and 2,660 offered at the coin flip. Why liquidity decides whether you can trade at all covers what happens to a book like this when news lands: the offers just above the bid vanish, the next real offer sits more than 20 cents higher, and the "price" jumps on nobody's opinion changing.
The third is the thing this whole page is about. The price is the crowd's probability of the contract's event, and the contract's event is a second announcement. A trader who buys at 39 because the Navier-Stokes proof looks solid has bought the wrong question. This board joins the daily Kalshi hub from here, where every price is quoted as a bid and an ask side by side, because on a book this thin that is the only honest way to quote one.
More on this: OpenAI AGI Odds: Altman Moved His Date. The Market Didn't. · OpenAI IPO: The CFO Named The Year. The Market Named Spring · Barcelona Vs Feyenoord Odds: Does The Favorite Survive The Fee? · The Longest Shot On This Board Might Be The Best-Kept Secret · 2026 Governor Races: Every Kalshi Board We Track
The Bottom Line
Kalshi built a contract that skips the fight everyone is having. The proof can be wrong, disputed, or never published in a journal, and none of it settles a single share; the Clay Institute's two-year clock will still be running when the far contract expires. The only referee is OpenAI's own blog, and the only question is whether the company posts a second announcement about one of the five problems left.
OpenAI has told us it pointed agents at all the open problems before pulling them onto Navier-Stokes. Whether that is evidence they are harder, or evidence of a process that now works, is the disagreement the mid-30s price is holding, and nothing on this page tells you which side of it to take.
The far contract's first buyer paid 98 cents for "again" on Tuesday evening. Then 24 hours later the same contract was 50 bid, 51 ask. The price of a repeat had already been argued down by half before most people had finished reading about the first one.
