On Friday, November 20, 2026, Anthropic mathematician Levent Alpöge published a statement claiming that he and the artificial intelligence system Claude have discovered a rigorous proof of the Riemann hypothesis. The announcement, released via Alpöge's personal research site alongside an enormous formal verification codebase, targets what is widely considered the holy grail of pure mathematics and the most famous of the Millennium Prize Problems.
First formulated in 1859 by Bernhard Riemann, the hypothesis posits that all nontrivial zeros of the Riemann zeta function have a real part equal to 1/2—a distribution rule that governs the hidden order of prime numbers. If verified by independent peers and automated theorem checkers, Alpöge's claim would resolve a 167-year-old intellectual roadblock, fulfill the criteria for a $1 million bounty from the Clay Mathematics Institute, and establish artificial intelligence as a primary engine of higher mathematical reasoning.
The Millennium Breakthrough: What Alpöge and Claude Claim
Alpöge, a junior fellow in the Harvard Society of Fellows and a member of the technical staff at Anthropic, published a four-page summary titled "Remark" asserting that he and Claude uncovered an end-to-end logical argument proving the conjecture. Rather than limiting the scope to Riemann's original statement, Alpöge stated that their approach smoothly generalizes to resolve the Generalized Riemann Hypothesis (GRH) for Dirichlet L-functions, which governs primes distributed across arithmetic progressions.
In his memo, Alpöge described a symbiotic division of labor. He provided high-level geometric direction and framed the problem using analogies derived from simpler toy models, while Claude contributed key technical maneuvers, constructed bridging arguments, and navigated deep literature on automorphic forms for higher-rank groups.
Key pillars of the announced proof include:
- A Reformulated Analytic Framework: Alpöge directed Claude to recast the behavior of the zeta function through weighted prime sums known as von Mangoldt sums.
- Higher-Rank Theta Lifts: Claude orchestrated a chain of constructions using automorphic forms to enforce precise square-root error bounds.
- Generalization to Dirichlet L-Functions: The methodology reportedly extends without major restructuring to settle GRH across multiple algebraic fields.
- Formal Verification in Lean 4: Every mathematical deductive step was translated into Lean 4 code to permit deterministic, computer-executed validation.
Watch: Claude VS Riemann?
Video: Amine on YouTube
Inside the Archive: 175,000 Files and a 250,000-Page Shadow Proof
Photo: The Verge (source)
To sidestep common criticisms of AI-generated "hallucinations" and dense mathematical prose, Alpöge avoided submitting a conventional mathematical preprint. In a footnote, he revealed that the raw, informal natural-language write-up generated by Claude exceeded 250,000 pages—a volume he deemed an "unreadable travesty of slop" that would overwhelm human reviewers.
Instead, Alpöge released a machine-checked codebase packaged into 14 compressed archive files totaling approximately 239 MB. Once decompressed, the project contains 175,187 source files formatted for the Lean 4 interactive theorem prover. By delivering the proof as code, Alpöge has invited the global mathematical and computer science community to verify the work computationally rather than relying on human endurance.
| Attribute | Alpöge & Claude Riemann Proof (2026) | OpenAI October Math Dump (2026) | Anthropic August Zero-Bound Milestone (2026) |
|---|---|---|---|
| Target Problem | Riemann Hypothesis & GRH | 372 topics (Quasi-Riemann, Hodge) | Zeta function critical line bound |
| Claimed Result | Complete proof (100% of zeros) | Zero-free region Re(s) > 7/8 | Increased bound from 41.6% to 67.2% |
| Format | Lean 4 formalization + 4-page note | 722 PDF manuscripts + partial Lean | Internal research write-up + note |
| File Size / Volume | 14 archives, ~239 MB, 175,187 files | 722 standalone papers on GitHub | ~31 million generated tokens |
| Human Guidance | High-level direction by Levent Alpöge | Semi-autonomous agent swarms | Prompt-driven iteration (Jared Sumner) |
| Prize Eligibility | Pending 2-year peer review / audit | Did not fully solve Millennium targets | Partial bound (non-qualifying) |
The Road to the Riemann Claim: A Year of AI Math Escalation
Friday's announcement is not an isolated event; it represents the culmination of a fierce, year-long arms race among frontier AI laboratories tackling open mathematical conjectures.
Alpöge first shook academic circles in July 2026 when he used an Anthropic model to discover an explicit counterexample to the 87-year-old Jacobian conjecture in three dimensions. One month later, he paired with mathematician John-Paul Smith to disprove the 1924 Carathéodory conjecture using Claude. Anthropic followed that achievement in mid-August by demonstrating that an internal research build of Claude could coordinate 60 sub-agents across 31 million tokens to raise the known proportion of zeta zeros on the critical line from 41.6% to 67.2%.
The competitive landscape turned contentious in September 2026. After rumors circulated that Alpöge and New York University mathematician Tristan Buckmaster were closing in on fluid dynamics problems, OpenAI deployed a cluster of roughly 10,000 autonomous agents spending millions in compute over 88 hours to claim a blowup singularity in the Navier-Stokes equations—sparking heated public debates over data access, prompting ethics, and academic credit.
Anthropic vs. OpenAI: Two Radically Different Philosophies for AI Mathematics
This latest development highlights fundamentally diverging strategies between OpenAI and Anthropic regarding AI research in the exact sciences.
OpenAI made waves in October 2026 with its "mathocalypse" dump—releasing 722 AI-generated manuscripts spanning algebra, probability, and topology, including a weaker "quasi-Riemann hypothesis" that proved a zero-free boundary rather than the hypothesis itself. OpenAI emphasized autonomous reasoning volume, averaging three hours of model deliberation per problem across 4,000 attempted questions.
In contrast, Anthropic's approach with Alpöge centers on deep human-machine symbiosis. Rather than flooding repositories with hundreds of machine-written PDF papers, Alpöge used Claude as an interactive reasoning partner to probe specific technical barriers in analytic number theory. By constraining the final public artifact to deterministic Lean 4 formal code, Anthropic aims to insulate its findings from the ambiguities of synthetic prose.
The Math Community’s Verdict: Skepticism, Fascination, and the Lean Audit
Reactions from number theorists, computer scientists, and academic bodies have ranged from awe to measured skepticism. While Lean formalization theoretically eliminates logical missteps, the software is only as sound as the axioms and definitions inputted into the environment.
Within hours of the download links appearing on alpo.ge, distributed teams began setting up verification pipelines. Initial assessments confirmed that the archive's cryptographic hashes match and that Alpöge's core theorem statements can be successfully expressed against Lean's mathematical library (mathlib). However, completing a full kernel-level compilation across 175,187 source files is expected to require significant server runtime.
Outside experts emphasize several critical audit requirements:
- Checking for "Sorry" Escapes: In Lean, an incomplete sub-proof can be bypassed using placeholder commands (such as
sorry). Independent analysts must certify that no unproven shortcuts exist within the 239 MB dependency tree. - Axiomatic Consistency: Auditors must verify that Alpöge and Claude did not introduce non-standard or contradictory axioms that might accidentally prove the conjecture via logical triviality.
- Faithfulness of Definitions: The mathematical formulation of the Riemann zeta function encoded in the code must precisely match Bernhard Riemann's established definition.
Alpöge himself struck a remarkably cautious tone in his note, admitting that he presented the Lean package before fully mastering the broader mathematical terrain Claude navigated. He acknowledged that the proof likely contains inefficient mathematical detours and predicted that human researchers will eventually compress the argument into much shorter, clearer proofs.
What a Confirmed Proof Would Mean for Math and Modern Computing
If the Lean codebase compiles cleanly and its mathematical foundations remain uncompromised, the impact will reverberate across multiple scientific disciplines.
In pure mathematics, hundreds of prominent theorems in analytic number theory and arithmetic geometry are conditional—meaning they begin with the preface, "Assuming the Riemann hypothesis is true." A definitive proof instantly solidifies those conditional proofs as objective mathematical facts.
In applied science and cybersecurity, the Riemann hypothesis governs the error bounds surrounding the prime counting function $\pi(x)$. Because modern public-key cryptography algorithms (such as RSA and elliptic curve variants) depend heavily on prime generation, pseudo-randomness, and factorization properties, confirming the exact distribution of primes provides cryptographers with rigorous theoretical bounds on the predictability of large prime sequences.
What Comes Next: The Two-Year Clock and Unanswered Questions
The Clay Mathematics Institute will not be cutting a $1 million check anytime soon. Under the official rules of the Millennium Prize Problems, a solution posted to a personal website or software repository does not qualify for the award.
To be considered by the CMI advisory committee, the proposed proof must be published in a recognized, peer-reviewed mathematical journal and withstand at least two full years of open scrutiny by the international mathematical community without being refuted. Furthermore, the committee will have to evaluate how to handle authorship, as current prize bylaws were written exclusively for human researchers.
In the coming weeks, attention will center on automated compilation nodes running the Lean 4 kernel, followed by human workshops aimed at reverse-engineering Claude's 175,000 code files into a concise, human-readable narrative. Whether Alpöge and Claude have truly climbed math's tallest mountain—or merely assembled a vast syntactic puzzle—will soon be decided by machine kernels and human mathematicians working in tandem.
FAQ
Did Claude solve the Riemann hypothesis by itself? No. The work was an active collaboration led by Anthropic mathematician Levent Alpöge, who provided conceptual framing, strategy, and problem-structuring while relying on Claude for technical derivations and Lean 4 code translation.
Is the Riemann hypothesis officially solved today? Not yet. While Alpöge has posted a completed Lean 4 formalization, the codebase must undergo independent verification to ensure it compiles without logical shortcuts, non-standard axioms, or faulty definitions.
Why did Alpöge publish code instead of a traditional research paper? Alpöge noted that an informal natural-language write-up generated by Claude exceeded 250,000 pages, which would be impossible for human mathematicians to review. He chose to release a machine-verifiable Lean 4 formal proof instead.
Will Anthropic or Alpöge win the $1 million Millennium Prize immediately? No. The Clay Mathematics Institute requires any candidate solution to be published in a qualifying peer-reviewed journal and to remain widely accepted across the global mathematics community for at least two years.





