How many more Millennium Prize Problems will AI solve in 2026?

The market is asking for an extremely specific and demanding event: not just major progress, but a qualifying announcement from one of a limited set of eligible AI labs, stating that one of five Millennium Prize Problems has been solved. As of mid-September 2026, there are only a few months left in the year, and the standard for success is not merely finding an interesting proof sketch or a promising theorem. The announcement must credibly present the problem as solved, and it must come from the right kind of organization. That combination makes the event materially harder than a generic AI breakthrough headline, so the default expectation should still be zero solved problems in 2026.

Arguments for Yes are not trivial, though. Frontier AI systems have already been pushing into advanced mathematics, formal reasoning, and theorem-proving workflows, and the incentive structure for leading labs strongly favors publicizing a landmark result if one exists. Because the eligible set includes many top-tier AI organizations, it does not take a broad industry consensus for the event to resolve Yes; it only takes one lab, or one official representative, making a sufficiently direct announcement. If a model-assisted breakthrough on one of the easier-to-attack Millennium problems were close to complete, the current period could plausibly be when a lab tries to secure first-mover advantage and public recognition before year-end.

Arguments against Yes remain stronger. The Millennium Prize Problems are famous precisely because they have resisted sustained effort for decades, and AI systems, while improving rapidly, still face major weaknesses in originality, long-horizon proof search, and verification of deep mathematical arguments. Even when models contribute meaningfully, the path from useful partial result to a universally credible full solution is long, and labs are usually cautious about making absolute claims unless the proof has been heavily checked. With only a short time remaining in 2026, the chance that a brand-new AI-driven proof of one of these problems emerges, is verified internally, and is then boldly announced as solved is still low.

The current market price implies a 42.5 percent chance of one or more solutions, which feels somewhat high relative to the combination of time remaining and the historical difficulty of these problems. A nonzero chance is justified because AI progress has been unusually fast and because the event definition depends on announcements rather than formal community acceptance, but the most likely outcome is still that no eligible lab will make a qualifying claim in time. The balance of evidence supports a meaningful but not dominant risk of a breakthrough, which is why the probability of zero solved problems is higher than the market-implied price suggests.