AI Math Conjecture Bet Hits 98% on Manifold: What Is Driving the Odds

An AI math conjecture market on Manifold just hit 98% YES, and the timing is not a coincidence. Traders are pricing near-certainty that an AI system will solve an important mathematical conjecture before January 1, 2030, after two AI labs posted results in May 2026 that leading mathematicians called genuine breakthroughs. Momentum has been building for months: a related Manifold forecast on the same question was already pricing 79% odds back in May 2026, just after those results broke, before climbing toward today’s near-lock. As of July 12, 2026, the market implies a 98% probability of YES, and a large share of its lifetime volume changed hands in a single recent session.

What the market is actually asking

The market lives on Manifold Markets, a play-money prediction platform, under the question “Will an AI solve any important mathematical conjecture before January 1st, 2030?” It was created by user MatthewBarnett. The resolution bar is deliberately high. To count as “important,” a conjecture must appear on a recognized list such as the Clay Mathematics Institute problems, Open Problem Garden, or Richard K. Guy’s Unsolved Problems in Number Theory — the same roster of Millennium Problems that Polymarket drew on when it launched a 14-outcome market in September 2026 asking which specific conjecture AI would crack first, pricing the Hodge conjecture at 22% odds.

Beyond the pedigree of the problem, the criteria require publication in a peer-reviewed mathematics journal, a proof judged correct by reviewers, no retraction within six months, and clear attribution to an AI system that autonomously discovered the primary breakthrough. That last clause matters: assistance is not enough, and the discovery has to be summarizable in a compact write-up. In other words, the market is not asking whether AI can help mathematicians. It is asking whether AI can do the central work itself and survive peer review.

The odds and what they imply

As of July 12, 2026, YES trades at 98% implied probability on Manifold. Implied probability is simply the market price read as a percentage: a 98-cent YES share reflects a crowd view that the event is nearly certain to happen before the 2030 deadline. If you are new to reading these numbers, our explainer on implied probability walks through how prices map to odds.

The move is notable not only for its level but for its speed. Manifold data shows the market has drawn hundreds of traders, and roughly a quarter of all lifetime volume traded in the most recent 24-hour window covered by the source, with hundreds of unique participants active. That is the signature of a market reacting to news rather than drifting. The remaining 2% of doubt is mostly about process, not capability: whether a specific AI result clears the peer-review and attribution hurdles cleanly, and who has authority to resolve the question given the creator has been inactive.

What is driving the move

Two developments in May 2026 reset expectations. On May 20, 2026, OpenAI said one of its general reasoning models disproved the Erdos unit distance conjecture, an 80-year-old problem in discrete geometry that Paul Erdos posed in 1946. The planar unit distance problem asks for the maximum number of point pairs exactly distance 1 apart among n points in the plane. The long-standing belief was that grid-like arrangements were essentially optimal. The model instead produced an infinite family of point arrangements that beats that bound, contradicting the conjectured limit.

One day later, on May 21, 2026, Google DeepMind published work describing an agentic proof system that autonomously solved 9 of 353 open Erdos problems and proved dozens of additional sequence conjectures, with results formalized and machine-verified in Lean 4. Machine verification is a meaningful detail, because a Lean proof can be checked by a compiler in seconds rather than waiting months for human referees.

The people and organizations involved

The names behind the results carry weight. Princeton mathematician Will Sawin refined the OpenAI geometry result, showing the improvement over the old bound can be made explicit. Fields Medal winner Tim Gowers described the work as a milestone for AI in mathematics, and other senior mathematicians have publicly endorsed the DeepMind proofs as real contributions rather than pattern matching. That endorsement from established researchers is part of why traders moved so decisively: the results were validated by the same community that would ultimately referee any resolution.

The two labs, OpenAI and Google DeepMind, announced within a day of each other, which amplified the signal. When two independent groups reach comparable milestones on the same weekend, it reads to the market less like a fluke and more like a capability threshold being crossed. The competitive framing extends beyond Manifold: a separate Polymarket contract on which lab will field the best math AI model has traded on more than $7 million in volume, with Anthropic priced between 89% and 99% as of July 2026, underscoring how closely bettors are now tracking the three-way race among OpenAI, DeepMind, and Anthropic. Polymarket has since broadened that appetite: by September 2026 it had listed a separate 14-outcome market on which specific conjecture AI will resolve first, set to settle by December 31, 2026.

The subtlety traders are pricing

Here is the tension the market has to resolve. Both May results are striking, but neither obviously satisfies every clause of the resolution criteria yet. The OpenAI result disproved a conjecture rather than proving one, which still counts as resolving it, but the criteria also demand peer-reviewed journal publication and clean AI attribution. The DeepMind proofs are machine-verified, yet formal verification in Lean is not the same as acceptance in a journal with a stated impact factor. So the 98% is best read as confidence that at least one such result will clear the full bar well before 2030, not a claim that the box is already checked.

This is a recurring theme in prediction markets: the gap between a real-world event happening and a market formally resolving. Definitions, timelines, and evidence standards can keep a market short of 100% even when the underlying trend looks settled — consider that Polymarket’s own 14-outcome conjecture market, set to resolve by December 31, 2026, prices the Hodge conjecture at just 22% even as Manifold’s broader yes-or-no question sits near-certain, a reminder that which conjecture and by when can matter as much as whether. Our primer on how markets resolve covers why wording is often as important as the news.

The broader context for a tech and finance reader

AI progress in mathematics has accelerated on a visible curve. In 2024, DeepMind’s AlphaProof and AlphaGeometry systems reached silver-medal standard at the International Mathematical Olympiad. In 2025, both DeepMind’s Gemini Deep Think and an OpenAI model reached gold-medal scores, solving five of six problems, and by December 2025 DeepSeekMath-V2 scored 118 out of 120 on the Putnam 2024 exam while hitting gold-level performance on IMO 2025 problems. Formal theorem-proving benchmarks widened the gap between frontier and open models further still: in March 2026, Meituan’s LongCat-Flash-Prover reached 97.1% on the MiniF2F-Test benchmark, and two months later, in May 2026, Claude Opus 4.5 posted 70.1% pointwise accuracy on a separate formal proof-search benchmark, well ahead of Qwen2.5-72B-Instruct’s 39.8% and Godel-Prover-V2-32B’s 36.4% on the same test. The 2026 Erdos results push past competition problems into open research questions that had resisted human effort for decades.

For finance readers, the relevance is not abstract. Automated mathematical discovery feeds directly into optimization, cryptography, and algorithm design, all of which have commercial value, and capital has followed: Renaissance Philanthropy and XTX Markets committed $18 million to an AI for Math Fund in September 2025, with individual grants of up to $1 million, while startups built around proof automation have raised at a rapid clip. Axiom went from a $64 million seed at a $300 million valuation in September 2025, led by B Capital, to a $200 million Series A at a $1.6 billion post-money valuation by March 2026, led by Menlo Ventures, and Harmonic AI reached a $1.45 billion valuation after a $120 million Series C led by Ribbit Capital in December 2025. It also sharpens the debate about how quickly frontier AI capabilities are compounding, a question that sits underneath valuations across the sector. The same appetite is visible in adjacent markets, such as the one tracking a perfect AI score at the 2026 IMO.

Timeline of the AI math milestones behind the bet

What to watch next

The near-term question is procedural. Watch for the OpenAI and DeepMind results to move from arXiv preprints and formal proofs into peer-reviewed journals, since that is the clause the market explicitly requires. Watch also for how the market handles resolution authority given its inactive creator, a governance wrinkle that has surfaced in the comments. If a headline result formally publishes and is cleanly attributed to an AI system, expect the market to converge toward 99% or resolve YES outright.

On the capability side, the next competitive benchmarks are the harder open problems. DeepMind and others have signaled interest in longer-horizon targets, and each new peer-reviewed result narrows the space for skeptics. For readers tracking the mechanics of these markets, our prediction market guide and overview of how these markets work are useful background.

Frequently asked questions

Does the market mean AI has already solved a major conjecture? Not formally. Traders are pricing near-certainty that a qualifying result will publish and clear peer review before 2030. The May 2026 results are strong evidence, but the market’s resolution criteria demand journal publication and clean AI attribution.

What counts as an “important” conjecture here? The problem must appear on a recognized list such as the Clay Mathematics Institute problems, Open Problem Garden, or Richard K. Guy’s Unsolved Problems in Number Theory, and the proof must be published and reviewed.

Is disproving a conjecture the same as solving it? For this market, resolving a conjecture either way counts. OpenAI disproved the Erdos unit distance conjecture, which settles the open question even though it shows the original claim was false.

Can I bet real money on this? Manifold is a play-money platform, so this specific market does not involve cash stakes. Regulated real-money venues exist for other questions, with different rules and restrictions.

The Bottom Line

Sources

• TechCrunch — OpenAI claims it solved an 80-year-old math problem, for real this time
• Scientific American — AI just solved an 80-year-old Erdos problem, and mathematicians are amazed
• The Conversation — An AI solution to an 80-year-old problem has shocked mathematicians
• arXiv — Advancing mathematics research with AI-driven formal proof search
• Google DeepMind — Gemini Deep Think achieves gold-medal standard at the IMO
• Manifold Markets — Will an AI solve any important mathematical conjecture before January 1st, 2030?

Prediction markets carry risk and are not investment or betting advice. Market availability is restricted by jurisdiction: Polymarket is not available to US persons, while Kalshi is a CFTC-regulated US exchange. Manifold operates as a play-money platform. Participation is limited to those 18+ or 21+ as applicable in your area. If gambling is a problem for you or someone you know, call 1-800-GAMBLER for confidential help.

Morgan Reeves

Prediction Markets Editor

Morgan Reeves covers prediction markets, sports analytics, and decentralized forecasting platforms. Background in quantitative finance and behavioral economics. Tracks Kalshi, Polymarket, PredictIt, and emerging US-regulated event-contract platforms since 2022.

View all articles