Superhuman math AI before 2030: Manifold bettors push the odds to 84%. A long-running Manifold market asking whether AI will reach superhuman mathematical problem solving before 2030, assuming no AGI arrives first, sits at an 84% implied probability of YES as of September 22, 2026. The jump follows two weeks of high-profile claims from OpenAI, Anthropic and NEAR AI, and a public rebuke from 25 Fields Medalists. Here is what the money is saying and what actually happened.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What the market is measuring
The market, created by Daniel Reeves (@dreev), asks a deceptively simple question: will AI achieve “superhuman mathematical problem solving before 2030, assuming no AGI yet?” The bar is high. Under the stated resolution criteria, an AI must be able to solve essentially any problem a team of Fields Medalists could solve when it is written on a single sheet of paper, max out benchmarks such as FrontierMath, deliver gold-medal International Mathematical Olympiad performance, ace the Putnam exam, and dominate humans on every metric including cost and speed. The question resolves around 2029, and it resolves N/A if artificial general intelligence, defined as the automation of essentially all remote human labor, arrives first.
That last clause matters for readers who follow the broader AI wagering scene. This is not a bet that machines will be smart at everything. It is a narrower bet that a specific, measurable form of mathematical superiority lands before the decade closes, and before the more sweeping AGI milestone that traders price separately in markets like our coverage of AGI within five years and the far more skeptical AGI before 2027 contract.
The odds as of September 22, 2026, and what they imply
As of September 22, 2026, the Manifold market reads 84% implied probability of YES. The source, Manifold’s dreev market, also showed top-tier 24-hour volume with 178 unique traders active, an unusually crowded book for a niche research question. Manifold runs on play-money mana rather than cash, so the number is best read as a crowd-sourced confidence signal rather than a regulated financial price. For a primer on how to translate a market percentage into a probability estimate, our explainer on reading Manifold odds covers the mechanics.
An 84% reading implies traders now see the “yes” outcome as the strong favorite, with roughly one-in-six odds against. That is a notable move for a proposition that, a year ago, most observers would have filed under speculative. The catalyst was not a single event but a compressed sequence of announcements in early September.
What is driving the move: OpenAI’s Navier-Stokes claim
The loudest catalyst arrived on September 8, 2026, when OpenAI said an unreleased internal model had produced a proof concerning finite-time blowup in a forced three-dimensional Navier-Stokes formulation. According to CNBC, the run used roughly 10,000 coordinated agents over about 88 hours, followed by an additional 17 hours of Lean formalization and verification. OpenAI’s own writeup describes the agents exchanging millions of messages and consuming on the order of hundreds of billions of output tokens across the full effort.
The important caveat, and the reason this did not simply resolve the debate, is that the result addresses the forced version of the equations, while the Clay Mathematics Institute’s Millennium Prize is defined on the unforced case. As Forkast reported, the work was not peer reviewed, was not accepted by the Clay Mathematics Institute, and drew a provenance dispute referencing mathematicians Tristan Buckmaster and Levent Alpoge. In other words, a technically valid proof of a nearby problem, not a settled Millennium Prize solution.
Anthropic, Fermat, and the formalization wave
Three days earlier, on September 5, 2026, Anthropic said Claude had completed the first fully formalized, machine-checked Lean proof of Fermat’s Last Theorem. Per Anthropic’s research post, the effort ran largely autonomously over about 11 days, produced roughly 13 million lines of Lean and proved around 29,500 intermediate theorems. Imperial College London mathematician Kevin Buzzard, who reviewed the work, called it an extraordinary autoformalization achievement.
Context is essential here. Fermat’s Last Theorem was proved by Andrew Wiles, with Richard Taylor, in the mid-1990s. What Claude produced is a translation of that known argument into a language a computer can mechanically check, not an independent discovery of new mathematics. That distinction is exactly what separates the high-probability and low-probability interpretations of the market.
NEAR AI and the collapsing cost of formal proof
The third September data point came from NEAR AI. On September 6, 2026, NEAR co-founder Alex Skidanov said an open-source Lean agent had solved all 672 problems in the PutnamBench benchmark for about $111 in total compute. CryptoBriefing reported that this undercut the next-cheapest known submission by roughly 250 times, with leading systems previously spending in the tens of thousands of dollars to clear the same suite.
The cost angle is why finance-minded observers took notice. If machine-verified theorem proving on competition-grade problems is becoming cheap and scalable, the economics of formal mathematics change quickly. The caveat is that a fixed benchmark can be gamed through memorization, contamination or dataset-specific engineering, so independent replication still matters.
The mathematicians pushing back
The countervailing signal arrived on September 11, 2026, when 25 Fields Medal winners, including Terence Tao, signed a declaration warning that the AI industry’s race to solve famous problems as benchmarks is harming the field. As Scientific American reported, the signatories argued that the goals of AI companies and the mathematical community are severely misaligned, and that rushed announcements leave too little time for careful writeups, verification and citation of prior work. The declaration proposed no specific rules or enforcement, but it framed the central tension precisely.
That tension is the crux for the market. A proof assistant can confirm that a formal derivation follows from its formal premises. It cannot, on its own, confirm that the formalized statement is the intended problem, that the encoding hides no loophole, or that the result is mathematically significant. Human experts still supply that judgment, and they are asking for time.
Why “superhuman” is doing a lot of work
The market’s 84% reading blends several outcomes that are not equally likely. It is worth separating them, because a tech or finance reader deciding what the number means should not conflate benchmark dominance with open-ended research ability.
| Capability | Recent evidence | How settled |
|---|---|---|
| Competition and benchmark performance | IMO gold-equivalent results; PutnamBench cleared for about $111 | Strong, at or above elite human level on fixed datasets |
| Formalizing known mathematics | Claude’s Lean proof of Fermat’s Last Theorem | Advancing fast, but formalizing existing results |
| Novel, important research result | OpenAI’s forced Navier-Stokes claim | Unverified announcement, disputed, not peer reviewed |
| Broad autonomous research across fields | No clear demonstration to date | Unresolved |
Read this way, the high-probability piece is that AI will beat elite humans on many fixed benchmarks and formalization tasks before 2030. The genuinely uncertain piece, and the one the market is really betting on, is whether machines will independently produce novel, important results that experts judge beyond unaided human capability, all without triggering the AGI clause that resolves the contract N/A.
What a tech and finance reader should watch next
Several concrete checkpoints will move this market from narrative to fact. First, whether OpenAI publicly specifies the exact Navier-Stokes theorem, releases the full proof, and lets its Lean formalization compile in a reproducible environment that independent mathematicians can inspect. Second, whether the Buckmaster and Alpoge provenance questions resolve cleanly. Third, whether the PutnamBench result survives independent audit for contamination. Fourth, whether the Fields Medalists’ declaration hardens into agreed norms for how AI math claims are announced and verified.
For traders, the parallel worth tracking is how tightly this contract stays coupled to the AGI markets. Because a sufficiently general system resolves this one N/A, the superhuman-math bet is partly a wager on sequencing: narrow mathematical superiority arriving before broad automation. That is a subtle position, and it is why the number sits at 84% rather than higher. Our ongoing coverage of the OpenAI math problem runs tracks the underlying capability curve.
Frequently asked questions
Does 84% mean AI has already achieved superhuman math? No. It reflects trader confidence on Manifold as of September 22, 2026 that the milestone lands before 2030. No result to date has been accepted by the mathematical community as a settled, novel breakthrough.
Did OpenAI solve the Navier-Stokes Millennium Prize? Not as defined by the Clay Mathematics Institute. OpenAI’s claim concerns a forced version of the equations, while the prize targets the unforced case. The work is unverified and disputed.
Is the Manifold market real money? Manifold uses play-money mana, so the percentage is a crowd confidence signal, not a regulated financial price. It is not a payout you can cash out like a CFTC-regulated contract.
Why does AGI matter to this market? The contract resolves N/A if AGI arrives first, so it is specifically a bet on narrow mathematical superiority preceding general automation.
The Bottom Line
Sources
- Manifold — Superhuman mathematical problem solving before 2030, assuming no AGI yet?
- CNBC — OpenAI claims to have solved the 90-year-old Navier-Stokes math problem in 88 hours
- Forkast — OpenAI’s 10,000-agent Navier-Stokes claim solves the wrong problem
- Anthropic — Formalizing Fermat’s Last Theorem
- CryptoBriefing — NEAR AI’s Lean agent solves every Putnam problem for $111
- Scientific American — 25 winners of math’s Nobel Prize decry the AI invasion of their discipline
Prediction markets carry risk and are not investment or betting advice. Odds move, and availability is restricted by jurisdiction: Polymarket is not available to US persons, while Kalshi is a CFTC-regulated US exchange; Manifold runs on play-money mana. Participation is limited to those who are 18 or 21 or older as applicable. If gambling is a problem for you or someone you know, call 1-800-GAMBLER for confidential help.
Related Coverage
- AI Beats Minecraft in Under 10 Minutes Before 2028? Manifold Bettors Put It at 70%
- Birch and Swinnerton-Dyer 'Solved Before 2045?' Odds Jump 38 Points on Kalshi After OpenAI's Math Claim
- 83% Odds: Manifold Bettors Say AI Will Play Any Game at Human Level by 2028
- Superconductor Bet Heats Up: Manifold Traders Give a Room-Temperature Breakthrough 15% Odds Before 2031
- Trump on US Currency? Manifold Traders Price 97% Odds His Face Appears Before 2028
