The offbeat market everyone is betting on: an AI program resolves the Riemann Hypothesis before 2035. A Manifold contract asking whether software will produce a machine-authored proof of mathematics’ most famous open problem is trading at 44% implied probability of YES as of August 2026, up from roughly 42% in July, and remains one of the busier questions on the platform. The odds have not moved because anyone proved anything — TechCrunch’s own August 2026 reporting confirmed the Clay Mathematics Institute’s $1,000,000 prize for the hypothesis remains unclaimed, a figure Anthropic’s own write-up on its zeta-bound milestone reaffirmed that same month. They have moved because a wave of genuine AI-and-math milestones, including an unreleased Anthropic Claude model’s advance on the underlying zeta-function bound reported in August 2026, a push that TechCrunch reported burned through roughly 31 million output tokens of compute that same month, plus a recurring set of viral hoaxes, has pushed the question back into public view.
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 actually asks
The contract lives on Manifold, a play-money forecasting platform, under the title “An AI program writes a correct proof that the Riemann hypothesis is either true, false, or undecidable (within the axioms of complex analysis) before 2035.” The resolution bar is deliberately high. The market description requires that any proof be vetted by qualified mathematicians to standards acceptable to a top journal such as the Annals of Mathematics, and that the argument originate primarily from the AI rather than from a human using a model as an assistant. In other words, a chatbot helping a professor fill in a lemma would not count. The software has to do the mathematics.
That distinction matters because it separates this bet from the more common story of 2026, in which humans and models work together. The market is pricing full machine autonomy on a research-level problem, not assisted proof-writing.
The odds, and what 42% implies
As of August 6, 2026, the Manifold market sits at 44% implied probability of YES, up from about 42% in July, with roughly 85 holders and several hundred trades logged on the contract. Tech Insider’s own tracking that same day counted zero peer-reviewed AI or human proofs of the Riemann Hypothesis in existence, a reminder that the price move reflects momentum rather than resolution. Because Manifold runs on its play-money currency rather than dollars, the figure is best read as a crowd forecast rather than a regulated financial price. For readers new to the mechanic, an implied probability is simply the market’s collective estimate that an event happens, expressed as a percentage. Our explainer on implied probability walks through how a price converts into odds.
A 44% reading is striking for a problem that, as Scientific American’s own August 2026 coverage noted, has resisted proof for 167 years. It does not mean traders expect a proof soon. It means a large minority of forecasters think there is a better-than-two-in-five chance that at some point in the next nine years, an AI system clears an extremely demanding bar, and the gap to even odds narrowed further after Anthropic’s August 2026 zeta-bound advance made headlines. That demanding bar shows up starkly on Riemann-Bench, a set of 25 expert-level problems tracked on arXiv, where frontier models were still scoring below 10% as of August 2026 — a reminder that headline momentum and actual capability are not the same thing. A separate framework posted to arXiv in April 2026 pushed the idea further still, proposing that large language models be set loose to discover entirely new major conjectures rather than just chase existing ones, including a family of Riemann-like problems modeled on the original hypothesis — an early sign researchers are starting to treat AI as a generator of open questions, not only a solver of them. The distance to 2035 is doing a lot of work here. For more on how these platforms translate belief into a tradable number, see our primer on how prediction markets work.
Why the market is moving now
There is no verified event in which any organization proved the Riemann Hypothesis outright — the hypothesis remains open. What has changed is the backdrop, and in August 2026 it shifted sharply: Anthropic disclosed that an unreleased Claude model raised the proportion of nontrivial zeta zeros proven to lie on the critical line from 41.6% to 67.2%, a 25.6-percentage-point jump detailed in Anthropic’s own blog post on August 10, 2026 and first aggregated by the industry newsletter Center Consulting, before TechCrunch published its own confirmation of the new 67.2% lower bound on August 11, 2026, with Europa Press following within days. The AI Canonical reported that the effort behind that jump wasn’t a single model run in isolation but roughly 60 subagents coordinated across two separate Claude Code sessions, testing some 650 distinct proof ideas as they ground through the search in parallel, a search that TechCrunch reported consumed roughly 31 million output tokens of compute that same month. Forbes reported that the model compressed roughly 37 years of equivalent human research effort into about 1.5 days of compute, a coincidence of numbers made sharper by Forbes’ separate note that the 41.6% bound Claude broke had itself stood as the field’s unchallenged record for that same 37-year stretch before August 2026. None of that proves the hypothesis itself, but that steady drumbeat of real results, rather than a single announcement, is the most-supported explanation for renewed interest in this specific contract.
The clearest recent signal came from Fields Medalist Terence Tao. In late 2025 and into 2026, Tao documented how large language models began contributing to the long-running list of Erdős problems. By his account, OpenAI’s GPT-5.2 Pro solved an open Erdős problem, number 728, largely on its own after some initial human feedback, one of the first cases where a model stayed faithful to the original question rather than retrieving a known answer. Tao has been careful about the framing, noting the win says more about speed than about difficulty, and that the broader future is “symbiotic” rather than a machine takeover.
The DeepMind and Gemini results
The other anchor is Google DeepMind. In a preprint dated January 12, 2026, DeepMind researchers working with mathematicians including Stanford’s Ravi Vakil used specialized systems built on Gemini to prove a novel theorem in algebraic geometry. Vakil described the model’s argument as no mere repackaging of existing proofs, calling it “the kind of insight I would have been proud to produce myself.” The American Mathematical Society’s leadership echoed that assessment.
Those are concrete, checkable milestones, and they are exactly what makes a market like this move. When a research reader sees a model produce an original proof in one field, the tail probability that software eventually cracks a Millennium Prize problem stops looking absurd. Anthropic’s Claude result comes closer than the others to touching the Riemann Hypothesis directly, since a proven bound on where zeta zeros sit is the same terrain as the hypothesis itself — even so, moving the confirmed share of zeros to about 67.25%, the refined figure Anthropic’s paper settled on and that Japanese tech outlet Gigazine covered in its own August 2026 write-up, a result reached only after the system burned through roughly 31 million output tokens of compute according to TechCrunch, is not a proof that all of them lie on the critical line — though the bound itself got an unusual degree of outside scrutiny, with mathematicians Conrey and Goldston verifying it and the result formalized in the Lean 4 proof assistant on August 11, 2026, according to the newsletter Temperature Zero. Algebraic geometry and an Erdős problem, by contrast, are not the same mountain at all.
The hoaxes feeding the hype
Not all of the attention is grounded in real work. The Riemann Hypothesis has become a magnet for viral claims. The most cited example is a 2024 post by xAI engineer Hieu Pham stating that Grok 3 had proved the hypothesis and that training was paused because the model had become “a danger to humanity.” Pham revealed hours later that it was a joke, and xAI colleagues leaned into the gag with escalating absurdities. The episode still circulates as if it were real.
Separately, social media posts through early and mid-2026 have claimed that a “quantum-AI” system produced a 500-page proof revealing a hidden pattern behind prime numbers. No such proof has been published or peer-reviewed. When forecasters see these claims recirculate, trading volume tends to tick up even though nothing has been proven. Sorting real milestones from noise is the entire skill here, much as it is when reading crowd bets on long-shot science questions.
Who and what is involved
- The Clay Mathematics Institute lists the Riemann Hypothesis among its Millennium Prize Problems, with a $1,000,000 reward for a correct solution — the same prize that a separate, older Manifold contract, “AI writes the proof to the Riemann Hypothesis by end of 2025,” still hinges on as of 2026, since both markets ultimately resolve against Clay’s criteria. That prize is part of why the question carries financial and cultural weight.
- Terence Tao (UCLA), the most prominent voice on AI’s real, limited contributions to research mathematics.
- Google DeepMind and collaborators including Ravi Vakil, behind the January 2026 Gemini algebraic-geometry result.
- OpenAI, whose GPT-5.2 Pro featured in Tao’s Erdős-problem write-ups.
- xAI and engineer Hieu Pham, the source of the recurring Grok proof hoax.
How this milestone compares
| Milestone | Status as of August 2026 | AI role |
|---|---|---|
| Riemann Hypothesis proof | Open, unproven for 167 years | None verified |
| Numerical zeros checked on critical line | Tens of trillions confirmed | Computational, not a proof |
| Erdős problem #728 | Reported solved, 2026 | Largely autonomous (GPT-5.2 Pro) |
| Algebraic-geometry theorem | Preprint, January 12, 2026 | Machine proof, human-vetted (Gemini) |
| “Quantum-AI” 500-page RH proof | Unverified viral claim | None; not published |
What a tech and finance reader should watch
The near-term tell is peer review. Several mathematicians expect 2026 to be the first year that AI-credited results clear review in major journals. If that happens for progressively harder problems, expect this market to drift higher regardless of whether the Riemann Hypothesis itself is touched. The signal to watch is not a headline claiming a proof. It is a formal publication in a venue like the Annals with an AI listed as a primary contributor. A parallel signal comes from benchmarking rather than publication: Surge AI has now run three rounds of its Riemann-Bench evaluation — September 2025, February 2026, and a third in June 2026 — and that latest round found frontier models still stuck below 10% on the benchmark’s 25 expert-level problems, a result that says more about real trajectory than any single viral claim. A third signal, this one financial rather than mathematical, is investor appetite: formal-math startup Harmonic closed a $120 million Series C in November 2025 at a $1.45 billion valuation, a bet by venture investors that AI-verified mathematics is now a fundable category rather than a research curiosity.
The second tell is resolution discipline. Because the contract demands a machine-primary proof vetted to journal standard, even a genuine human-AI collaboration on the Riemann Hypothesis could resolve this market NO. Traders who conflate “AI helped prove it” with “AI proved it” are pricing a different question than the one written. That gap between assisted and autonomous work is where the real disagreement sits. Markets tracking AI performance on hard reasoning benchmarks face the same definitional traps.
Frequently asked questions
Has any AI proved the Riemann Hypothesis? No. As of August 13, 2026, the hypothesis remains open, and Tech Insider’s count of peer-reviewed AI or human proofs still stood at zero as of August 6, 2026. An unreleased Anthropic Claude model did raise the share of nontrivial zeros proven to sit on the critical line to 67.25% that month, a genuine advance on the underlying math, but that is a bound, not a full proof, and no verified machine-authored proof of the hypothesis exists.
Did Grok 3 really prove it? No. That claim originated as a joke by an xAI engineer in 2024 and was walked back within hours. It keeps recirculating as if genuine.
What would it take to resolve the market YES? An AI-authored proof, vetted by qualified mathematicians to the standard of a top journal, with the argument originating primarily from the software rather than from a human using it as a tool.
Why is the market at 44% if nobody is close? The window runs to 2035, and Tech Insider’s tracking put the contract at that 44% level as of August 2026. Traders are pricing nine years of rapid progress in AI-assisted mathematics, not an imminent breakthrough. For context on reading crowd odds, see our guide comparing Kalshi and Polymarket.
The Bottom Line
Sources
- Manifold — AI program resolves Riemann Hypothesis before 2035 market
- Quanta Magazine — The AI Revolution in Math Has Arrived
- Scientific American — AI Uncovers Solutions to Erdos Problems
- The Decoder — Terence Tao on GPT-5.2 Pro cracking an Erdos problem
- OfficeChai — Google Gemini proves novel theorem in algebraic geometry
- AIbase — Grok 3 Riemann Hypothesis claim was a false alarm
- Forbes — Let’s Solve The Riemann Hypothesis
Prediction markets carry risk and are not investment or betting advice. Availability is restricted by jurisdiction: Polymarket is not available to US persons, and Kalshi is a CFTC-regulated US exchange. Manifold is a play-money platform. You must be of legal age (18+ or 21+ as applicable). If gambling is a problem for you or someone you know, call 1-800-GAMBLER for confidential help.
