An AI Proof of the Riemann Hypothesis Is Coming—If Mathematicians
Anthropic's mathematician says Claude helped solve the Riemann hypothesis. Whether the proof survives scrutiny matters more than who gets credit.
The Announcement That Timing Makes Complicated
Levent Alpöge, a mathematician at Anthropic, posted a note on his personal website last Friday claiming that Claude had helped him prove the Riemann hypothesis. Six weeks of work, he wrote. Claude did the heavy lifting near the finish line. He linked Lean code—a formal language designed to let computers check mathematical proofs line by line.
The proof remains unverified. No one has confirmed it yet. But the timing of the announcement has already drawn sharp attention: it came one day after Anthropic filed to go public on the Nasdaq. Alpöge acknowledged the awkwardness himself, calling it “horrible.” If the proof holds, it could reframe the entire conversation around AI-assisted mathematics. If it does not, it becomes another cautionary tale about hype outpacing rigor.
Either way, the mathematics community is watching closely. And it is watching with a mixture of hope and deep suspicion.
What Makes This Claim Different
Alpöge did not just publish a paper or release a press statement. He posted Lean code. That matters enormously. Lean is not a black-box reasoning engine. It is a formally verifiable system built specifically so that every step of a proof can be checked mechanically. If the code is correct—if the theorem follows from its axioms through valid inference rules—then no human doubt can touch it. Verification becomes a programming task rather than a philosophical one.
This is the central innovation that separates the current wave of AI-in-mathematics from earlier attempts. Claude may have guided Alpöge toward the right structure, suggested lemmas, spotted dead ends. But Lean would demand that every single implication be justified. A hallucinated step would not pass inspection. That is why the real story here is not whether Claude solved the Riemann hypothesis. It is whether the code works.
Independent researchers are currently running checks on Alpöge’s submission. Some early commentary from mathematicians on social media was encouraging but inconclusive. No peer-reviewed confirmation has appeared as of this writing.
Who Wins, Who Loses, What Happens Next
If the proof checks out, the immediate winners are Anthropic and the growing camp of researchers who argue that AI-generated mathematics will become mainstream. Alpöge, as lead author, would almost certainly share in any Millennium Prize award. The Clay Mathematics Institute has offered $1 million for a correct solution to the Riemann hypothesis—one of seven such problems identified in 2000. A verified AI-assisted proof would also set a powerful precedent for the other six unsolved problems.
Losers in this scenario are harder to name precisely but real none-the-less. Mathematicians who prefer purely human-driven proofs, like the Italian researcher Alessandro Masullo, who released an incomplete attempt last week after twelve years of work without relying on AI, find their approach framed as obsolete. So do the critics who argued that corporate-backed AI in mathematics was intellectual showcase rather than genuine progress.
But the most important question is whether the community trusts the result. Mathematics has always been skeptical of shortcuts to deep truths. The gap between “this looks correct” and “this is proven” has defined the discipline for millennia. Lean narrows that gap significantly, but it does not eliminate all concerns. Who reviewed the code? Did anyone audit Alpöge’s interaction logs with Claude? Was there any manipulation of the verification environment itself?
The OpenAI Shadow Looms Large
Alpöge’s claim arrives against the backdrop of a rival performance. In September, OpenAI announced that a non-public version of its tools had helped solve part of another Millennium Prize problem—using 10,000 instances of its AI systems at a reported cost of $15 million. The result struck many observers as less like genuine mathematical progress and more like brute-force industrial might displayed as insight.
James Robinson, a University of Warwick mathematician, called the OpenAI announcement “immature playground boasting writ large,” pointing to the environmental costs of running that scale of compute and the deeper concern that such displays privilege capital over cleverness. Alpöge was working on that same OpenAI-targeted problem when the announcement surfaced.
That context makes Alpöge’s Lean-based approach comparatively appealing. Where OpenAI relied on raw computational volume, Alpöge relies on formal verification. But the rivalry between Anthropic and OpenAI does not disappear when mathematicians examine the proof. If anything, it intensifies the incentive to scrutinize the work aggressively, since dismissing it strengthens one side of the competitive narrative.
Why Trust Remains the Real Hurdle
There is a deeper unease in the mathematical community that goes beyond individual proofs. The Riemann hypothesis has resisted solution since Bernhard Riemann proposed it in 1859. Nearly 170 years of the world’s finest mathematical minds have failed to crack it. Now an AI company appears to have done it—or at least half of it—in six weeks, during the same quarter that its parent company prepares to list on a major stock exchange.
Skepticism is not irrational. It is how the field protects itself. Every claimed proof of the Riemann hypothesis before this one has turned out to contain errors. The bar should remain high.
But the existence of Lean changes something fundamental about what counts as evidence. Human intuition will still matter—someone needs to understand whether the proof illuminates the structure of prime numbers or merely arranges them into a tautology. Yet mechanical verification is the closest thing mathematics has ever had to an objective standard. It reduces the role of authority, reputation, and rhetorical elegance to secondary factors.
What matters now is the next few months. Alpöge’s code must survive independent verification. Other mathematicians must reproduce the steps. The relationship between the AI generation process and the final Lean script must remain transparent enough for the community to judge whether Claude discovered the proof or merely dressed up an outline left behind by human intuition.
If all of that succeeds, mathematics gains a new kind of collaborator. If it fails—or if the proof turns out to rest on a flawed assumption hidden somewhere in the Lean code—the episode will reinforce the belief that AI-assisted math is useful for exploration but insufficient for conviction.
For now, the only responsible position is waiting. The code exists. The verification has not. When it comes, the mathematics community will know immediately whether this was a breakthrough or a mirage shaped by capital and timing.