The offbeat market everyone is betting on: will an AI solve an important math conjecture before 2030. A Manifold market asking exactly that question is priced at roughly 99% implied probability of YES as of August 21, 2026, one of the platform’s most heavily traded questions of the week, with 359 unique bettors active in the last 24 hours. The reason is not abstract: in the first two weeks of August, both OpenAI and Anthropic published machine-verified results on genuinely open mathematical problems, and forecasters repriced accordingly. The pattern held into September: a separate Manifold market asking whether a formalized proof of Fermat’s Last Theorem will exist by 2029 also reached roughly 99% implied YES, with 58% of that market’s entire lifetime trading volume changing hands in a single day in September 2026, a sign that betting interest in AI-verified mathematics is compounding rather than a one-off spike.
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 actually asking
The market, created by researcher Matthew Barnett on Manifold, poses a deceptively simple question: will an AI solve any important mathematical conjecture before January 1st, 2030? The resolution criteria matter here. Barnett defines “important” as a conjecture appearing on a recognized list, such as the Open Problem Garden, the Clay Mathematics Institute’s problems, or Richard K. Guy’s book Unsolved Problems in Number Theory. For a YES resolution, an AI system must fully solve one of these, with the solution published in a reputable mathematics journal.
That last clause, the journal requirement, is why the question is not already resolved despite a flurry of August announcements. AI systems have produced results on open problems, but the traditional peer-review pipeline moves in months and years, not days. The market is pricing the gap between “verifiable proof exists” and “the mathematics establishment has formally blessed it.”
The odds, and what 99% implies
As of August 21, 2026, the market trades near 99% implied probability of YES on Manifold. Manifold is a play-money platform, so this is a crowd forecast rather than a cash-settled price, and the number should be read as a confidence signal, not a payout. Related cash and play-money markets on adjacent questions sit lower and further out: forecasters have recently priced an AI-solved Millennium Prize Problem well below even odds for the mid-2020s, and a separate Manifold market on the Riemann Hypothesis by 2035 climbed from 42% to 44% between July and August 2026. Polymarket runs its own version of the question with a more granular structure: a 2026 AI-math market there carries 14 distinct outcomes as of September 2026 and had drawn $43,823 in trading volume by that point, while Polymarket’s broader math category had grown to roughly 500 listed markets, including several AI-model prediction markets, by October 2026, reflecting how differently each venue frames the same underlying uncertainty and how quickly interest in AI-verified mathematics is scaling across platforms. The 99% figure reflects the specific, looser wording of Barnett’s question (any important conjecture, on broad lists, by 2030) rather than a claim about the hardest problems in the field. For readers newer to these instruments, see our explainer on how prediction markets work and what implied probability actually means.
What is driving the move: OpenAI’s Astra
The single largest catalyst arrived on August 1, 2026, when OpenAI published a page titled “Ten advances in mathematics and theoretical computer science” and released a GitHub repository of Lean 4 proof files. The company said an unreleased internal model it calls Astra produced ten results on long-standing open problems, each accompanied by a machine-checkable Lean 4 certificate and documented in a manuscript reported at 249 pages.
The headline result, covered by Forbes and SiliconANGLE, is the first explicit construction of a non-sofic group, answering a question about whether every countable discrete group must be sofic that had stood since Mikhail Gromov’s work in 1999. Other results include a disproof of a rigidity conjecture associated with Fields Medalist Alain Connes in the theory of von Neumann algebras, three problems from Paul Erdos’s catalogue including problem 183 on multicolor Ramsey numbers, the first improvement to the general upper bound on high-dimensional sphere-packing density since 1978, and a parallel repetition theorem for two-player quantum games.
OpenAI estimated the token cost of finding all ten solutions at roughly $2,000 at its internal API rates. The repository’s “sorry” count, the Lean keyword that marks an unfinished proof step, was reported at zero, meaning every step across the ten formalizations passes the proof checker.
Why the verification method matters to forecasters
The reason this event moved a prediction market, rather than just generating headlines, is the verification path. A Lean 4 proof is machine-checkable: a skeptic can run the proof files on a laptop and confirm that each logical step follows, without trusting OpenAI’s narrative. That collapses one axis of doubt that normally surrounds AI claims about research.
The reaction from working mathematicians was cautious but notably serious. Fields Medalist Timothy Gowers commented positively while stressing the work is still being digested, and Thomas Bloom, the University of Manchester mathematician who curates the Erdos problems catalogue and who publicly debunked an earlier, weaker OpenAI claim in October 2025, called the Astra results “big news.” That last detail is important: the same researcher who had previously pushed back was, this time, taking the results seriously.
Anthropic and the Riemann Hypothesis
The second catalyst came on August 11, 2026, when TechCrunch reported that an unreleased Anthropic model made measurable progress on the Riemann Hypothesis, one of the Clay Millennium Problems. Anthropic’s own research writeup describes a research version of Claude increasing the proven share of Riemann zeta zeros known to lie on the critical line, with the company framing it as a substantial jump on a bound that had seen little human movement for decades.
The critical caveat, emphasized by Scientific American under the blunt headline “No, Anthropic’s AI didn’t just solve math’s hardest problem,” is that this does not prove the Riemann Hypothesis. It advances a related, quantitative bound. But for a market asking about “any important conjecture,” a credible partial result on a Millennium Problem, arriving ten days after Astra’s ten proofs, reinforced the direction of travel.
Timeline: the two weeks that repriced the market
| Date (2026) | Event | Why it mattered |
|---|---|---|
| Aug 1 | OpenAI publishes Astra’s ten results with Lean 4 proofs | First machine-verified AI solutions to decades-old open problems |
| Aug 2-3 | Coverage from Forbes, SiliconANGLE, TechTimes; Bloom and Gowers weigh in | Working mathematicians treat the non-sofic group result as genuine |
| Aug 11 | Anthropic reports Claude progress on Riemann zeta bound | Credible movement on a Millennium Problem, though not a proof |
| Aug 12-13 | Scientific American and others add caveats | Peer review and journal publication remain the open gap |
| Aug 21 | Manifold market near 99% YES, 359 bettors in 24h | Forecasters price the 2030 question as near-certain |
Who and what is involved
The named actors are concrete. On the AI side: OpenAI’s unreleased Astra model and Anthropic’s unreleased research version of Claude. On the mathematics side: Timothy Gowers and Thomas Bloom as public commentators, Alain Connes and Mikhail Gromov as the originators of the problems Astra addressed, and Paul Erdos and Richard K. Guy as sources of the curated problem lists the market references. The verification infrastructure is the Lean 4 proof assistant, an open-source tool that both companies used to formalize their claims. The market itself sits on Manifold, run by Manifold Markets, and was created by Matthew Barnett, a researcher known for AI-forecasting questions.
Broader context for a tech and finance reader
For markets and enterprise buyers, the signal is not that a chatbot did algebra. It is that a frontier model produced new, formally verified theorems for a compute cost reported around $2,000, and that independent parties could check the work without trusting the vendor. That math breakthrough did not land in isolation: it arrived in the middle of an unusually fast model-release cycle, with Google DeepMind shipping Gemini 3.1 Pro Preview on February 19, 2026, OpenAI following with GPT-5.4 on March 6, 2026, and DeepSeek pushing out DeepSeek V4 Pro 0423 on April 24, 2026, each pitched partly on reasoning and formal-math gains. That combination, novel output plus cheap, trustless verification, arriving amid a compressed release cadence across labs, is the pattern that turns a research demo into something a business can underwrite. It is also why prediction markets, which aggregate money and conviction rather than press releases, are a useful read on this story. Prediction platforms have been increasingly used as real-time gauges on everything from Fed rate cut odds to political outcomes, and the mechanics differ by venue. If you want the venue landscape, compare Kalshi versus Polymarket before assuming any given quote is a cash price.
What to watch next
Three things will decide whether the 99% forecast is vindicated on the market’s own terms. First, journal publication: Barnett’s resolution requires a solution to appear in a reputable mathematics journal, and refereeing timelines are the binding constraint, not the mathematics. Second, sustained expert consensus: informal praise from Gowers and Bloom is a strong early signal, but the field still needs time to fully vet the non-sofic group construction and the Connes-related disproof. Third, whether either lab, or a competitor, converts partial progress like Anthropic’s Riemann bound into a complete, refereed result before 2030. A fourth signal worth tracking is the adjacent Manifold market on a formalized Fermat’s Last Theorem proof by 2029, which was trading near 99% YES in September 2026 and saw 58% of its total volume trade in a single day, evidence that betting interest in AI-verified mathematics is building rather than cooling as the 2030 deadline nears. Any one of those clearing would likely resolve the market YES well before the deadline.
Frequently asked questions
Has an AI actually solved a Millennium Prize Problem? No. As of August 21, 2026, no Millennium Prize Problem, including the Riemann Hypothesis, has a published, peer-reviewed proof from an AI or a human. Anthropic’s work advanced a related bound; it did not prove the hypothesis.
What does the 99% number mean? It is the implied probability of YES on a Manifold play-money market as of August 21, 2026. Manifold uses virtual currency, so it reflects crowd confidence, not a cash payout, and applies specifically to the market’s broad wording about “any important conjecture” by 2030.
Are the Astra proofs confirmed correct? They are machine-checkable in Lean 4, meaning the logical steps pass an automated proof checker, and prominent mathematicians have reacted positively. They have not yet completed formal journal peer review.
Can I bet real money on this? Not on this specific Manifold market, which uses play money. Real-money event contracts exist on other platforms with different rules and jurisdictional limits: Polymarket, for instance, had grown to roughly 500 math-related markets by October 2026, and its own 14-outcome 2026 AI-solved-problems contract had traded $43,823 by September 2026; see the disclaimer below.
The Bottom Line
Sources
- Manifold — Will an AI solve any important mathematical conjecture before January 1st, 2030?
- Forbes — OpenAI’s Astra Solved Decades-Old Math Problems For $2,000
- SiliconANGLE — OpenAI’s Astra solves 10 long-open math problems and publishes the proofs
- TechCrunch — An unreleased Anthropic model made progress on one of math’s biggest unsolved problems
- Scientific American — No, Anthropic’s AI didn’t just solve math’s hardest problem
- Anthropic — Learning more about Claude’s mathematical capabilities
Prediction markets carry financial risk and are not investment or betting advice. Availability is restricted by jurisdiction: Polymarket is not available to US persons, while Kalshi is a CFTC-regulated US exchange. The Manifold market referenced here uses play money. Participation is limited to those 18+ or 21+ as applicable in your jurisdiction. If gambling is a problem for you or someone you know, call 1-800-GAMBLER for confidential help.
