A frontier AI model just spent roughly a day and a half on the Riemann hypothesis - one of the seven Clay Millennium Prize challenges, with a $1 million bounty for a complete proof - and came out the other side with a paper that two of its maker’s in-house mathematicians signed off on. The result is not a proof. The argument that mattered on Monday was not whether the math was right. It was whether the model should be on the author line at all.
That fight has a name now. It is the Leiden Declaration, a June 2026 statement signed by a group of mathematicians arguing that AI-generated proofs should not undermine the field’s standard that proofs are “attributable to specific authors who take credit for their discovery and assume responsibility for their correctness.” The Anthropic result is the first time that argument has collided with a named, dated, on-the-record research advance. The Leiden Declaration is the through-line from this week’s Riemann story to every AI-as-scientist claim that follows.
What the model actually did
The Riemann hypothesis concerns the distribution of prime numbers and the location of the zeros of the Riemann zeta function. A full proof has been open since 1859. The Anthropic model did not produce one. According to TechCrunch’s 11 August report, the model raised the lower bound for which the hypothesis holds - a real, incremental advance on the problem, not a solution.
The setup was unusual in a specific way. An Anthropic staff member with no significant mathematical training asked the model to “take a real stab” at the problem. The model then ran autonomously for about 1.5 days, coordinating 60 subagents, testing roughly 650 ideas, and emitting 31 million output tokens. Anthropic’s own breakdown: two subagents developed the key ideas, 13 contributed supporting ideas, 30 attempted and failed to develop new ones, 13 served as validators, and two helped write the paper. Two in-house mathematicians confirmed the result, and the work was formalised in the open-source Lean proof assistant - the same tool the wider mathematical community uses to verify machine-checked proofs.
Anthropic has not published the model. The result is real and the chain of custody is clean, but the model itself is one more entry in the frontier-lab pattern of running an unreleased system on a public-facing problem and publishing the output. OpenAI ran the same play on 1 August with an internal model the company calls Astra, generating ten new results in mathematics and theoretical computer science at an API cost of about $2,000 - including progress on a long-standing open group-theory question, a non-sofic-groups existence proof, and results across high-dimensional geometry, coding theory, quantum complexity, lattice cryptography, and extremal combinatorics. Noam Brown of OpenAI, asked about the Millennium Prize problems, told The Decoder that Astra had not solved any of them: “Sadly, no Millennium Prize Problems (yet).”
The practical point for readers: the frontier-lab consensus is now that the math tools work, the economics are negligible (OpenAI publicly put the cost of all ten Astra solutions at about $2,000 at API rates), and the unsolved questions are no longer technical. They are about credit, attribution, and what a “mathematical author” is.
The Leiden Declaration and the AI-author fight
The Leiden Declaration came out of a September 2025 workshop at the Lorentz Centre in Leiden, where mathematicians, historians, philosophers, and AI researchers sat down to write a joint statement. The version released in June 2026 makes three concrete claims: that proofs must be attributable to specific authors; that those authors must take responsibility for correctness; and that the existing scholarly record must be protected from a flood of unattributed AI-generated content. The Declaration is short, careful, and not anti-AI. It accepts that LLMs are useful tools. It draws the line at authorship.
Not every mathematician who attended the workshop signed. Fields Medalist Timothy Gowers was there and declined to sign, explaining his reasoning in a 26 July blog post. He agrees with most of the Declaration but calls parts of it “confident assertions and recommendations that I feel somewhat uncertain about.” His central thought experiment is whether to resist a future where AI autonomously produces verifiable proofs that LLMs can also explain. He lands on a softer line than the Declaration: “If we arrive at a world where mathematical theorems are no longer associated with mathematicians, maybe that won’t be any more problematic than the fact that stars aren’t named after astronomers and most aren’t named at all.”
Gowers’s harder concern is the second-order effect. If junior mathematicians stop spending the years it takes to build the expertise to read a frontier proof, the discipline could end up with vast archives of correct machine-checked work and no community capable of digesting it. His proposed fix is to shift credit away from whoever prompted the model and toward whoever makes the effort to digest the AI’s output and explain it - “more like the gratitude that one feels already for somebody who writes a beautiful textbook.” The Leiden Declaration wants authorship defended. Gowers wants authorship redistributed.
The Anthropic Riemann paper now sits in the middle of that argument. Two in-house mathematicians confirmed the result. Sixty subagents did the work. The question Anthropic has not answered publicly is which name(s) appear on the paper, and in what order. If the byline reads “Anthropic’s unreleased model” the Leiden Declaration loses its cleanest test case. If the byline reads “two in-house mathematicians” and the model is acknowledged in a footnote, the Anthropic model becomes the first publicly-named LLM to land a co-authorship credit on a Millennium Prize-class paper. Neither outcome has been disclosed yet.
What This Means
The credit fight is downstream of a quieter one. The 11 August Nature report on QED Science, a Tel Aviv start-up whose tool claims to identify the top 1% of bioRxiv preprints, sits on the same question from a different angle: should an AI system be allowed to evaluate scientific work? QED Science says yes, and the platform reports it is in use at more than 10,000 laboratories across 1,500 institutions in over 70 countries. Bibliometricians are not convinced; Nature quotes researchers worried that an AI-grade prestige badge will reinforce the metric culture the field has been trying to dismantle for a decade. The Leiden Declaration is the same argument, applied to the authorship line rather than the prestige line.
For a privacy-and-local-AI audience the through-line is what gets lost in the headline. The Anthropic and OpenAI results are technically real, economically cheap, and not yet reproducible by anyone outside the labs that produced them. The Leiden Declaration, Gowers’s counter-proposal, and the QED Science dispute are about the part of the scientific process the labs do not control - who counts as a scientist, who gets credit, who evaluates the work. That is the policy question the next year of AI-meets-science coverage will track, not whether the next model produces a longer proof.
The Bottom Line
Anthropic’s Riemann result is real but not a proof. The interesting fight is over the author line, and the math world is split between the Leiden Declaration’s defenders of human authorship and Gowers’s case for redistributing credit to whoever explains the AI’s work. Both sides agree on what the model did. They disagree on what to call the person who made it happen.