
On November 20, 2026, mathematician Levent Alpöge, working at Anthropic, publicly claimed to have completed a proof of the Riemann hypothesis with Claude as his primary tool. The manuscript is on a preprint server but has not passed peer review. If it holds, it would be the first resolution of a problem posed in 1859 and unsolved for 167 years, and the first candidate result from a general-purpose model on an open mathematics problem with no known route. The conclusion: AI has moved from
The timeline is concrete. On November 20, Alpöge posted a technical manuscript on a preprint server and listed Claude as the main instrument of the work. Alpöge has published previously in number theory and arithmetic geometry. By his own account the research ran in three stages: using Claude to map existing literature and recorded counterexamples around the distribution of zeros of the Riemann zeta function, then generating a proof skeleton and a dependency graph of lemmas, and finally checking the critical inequality estimates by hand. The Riemann hypothesis states that every non-trivial zero of the zeta function has real part one-half. The Clay Mathematics Institute listed it among seven Millennium Prize Problems in 2000, with a $1 million award. The manuscript's current state: the full written argument is public, the Lean formalization covers only part of it, no journal submission
Why this lands in 2026: progress in AI-for-mathematics has followed one clear path over three years. DeepMind's AlphaProof translated natural-language problems into Lean goals and then searched, solving four of six problems at the 2024 International Mathematical Olympiad; AlphaGeometry reached near-gold-medal level in plane geometry in the same period. Together they established a paradigm: replace unverifiable free generation with a state search that a machine can reject step by step. Meanwhile the Lean community and the mathlib lemma library kept expanding, turning formal checking from a specialist tool into callable infrastructure, and reasoning-oriented models became more stable on long
Liu Gong's take: this manuscript is very likely not a flawless final draft, but it compresses the timeline for AI participating in core mathematical proof from a decade to two or three years. Liu Gong offers a checkable forecast — within the next six months, if the team pushes the argument to full Lean formalization, at least one lemma will be found to need substantive repair, while the overall skeleton will hold.
From a technical standpoint, the Riemann hypothesis is hard for structural rather than computational reasons: it demands a chain of reasoning linking zeta zeros to prime distribution that nobody has walked before. Large models are strong at cross-literature association and lemma recombination, and weak at keeping intermediate steps stable over long chains, where one bad step invalidates everything after it. What matters, then, is not a model writing a proof in one pass but the loop of a human setting direction, a model searching at lemma level, and Lean rejecting step by step. AlphaProof solved four of six IMO 2024 problems through exactly that loop, not through free generation. A general model plugged into the same loop gains low migration cost across fields and loses the search-space control a dedicated prover provides.
On the business side, Anthropic gains a brand premium for trustworthy reasoning while carrying the risk that an unverified announcement rebounds. The dedicated-system route sinks cost into infrastructure — a formal environment and a custom search engine. The general-model route spreads cost across API calls, with unit cost falling at scale. That makes the former suited to small well-defined targets and the latter to large fuzzy ones. The Riemann hypothesis belongs to the second category.
For the industry, the pressure lands on review institutions. Preprints plus Lean checking could form a fast-track far quicker than journals, speeding priority races among mathematicians and making them messier. Liu Gong's experience in industrial visual inspection is that every verdict a model issues must trace back to specific pixels and criteria, or a production line will not use it. Formal checking of a proof and traceability in inspection are the same problem: replace trusting the output with verifying the process. The real value of this announcement lies less in whether it succeeded at once than in forcing mathematics to standardize how it verifies.
- The Riemann hypothesis was posed in 1859 and has stood for 167 years as of 2026; it is the only problem on Hilbert's 1900 list that also appears among the Millennium Prize Problems.
- The Clay Mathematics Institute set seven Millennium Prize Problems in 2000 at $1 million each; only the Poincaré conjecture has been confirmed solved, by Perelman in 2002-2003, with the award made in
- – DeepMind's AlphaProof solved four of six problems at the 2024 International Mathematical Olympiad, a silver-medal level, with every solution verified in Lean.
- Compared with dedicated provers, a general model needs no rebuilt formal environment per domain, but under equal compute it constrains the
For the industry, the first three months bring a jump in demand for formal verification: the Lean community will absorb a wave of new scripts from number theory, and journals and preprint servers will be pressed to define disclosure rules for AI-assisted proofs. Within a year, if the core skeleton survives review, the division of labour in mathematical research will shift — literature mapping and lemma search move to models at scale, while human researchers spend more time on problem selection and direction.
For ordinary users, the short term brings another round of marketing language, with AI proving a century-old problem becoming a generic talking point; the risk is treating an
Questions worth exploring:
- Whether formal verification becomes the de facto standard for AI mathematics, and the faster review and messier priority races it brings
- General models with human experts versus dedicated symbolic systems: which route suits open mathematics better
- A million-dollar prize and academic credit: how responsibility and honour are allocated when AI takes part in a proof
Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.
This item comes from The Verge AI (published 2026-11-20 16:34:13). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.
