Anthropic’s Unreleased Model Advances on the Riemann Hypothesis

Anthropic says an unreleased AI model has made meaningful progress on the Riemann hypothesis, the 150-year-old mathematics problem tied to the distribution of prime numbers and one of the most famous unresolved questions in the field. The result matters because it suggests frontier models may be getting better not only at solving routine problems, but at generating genuinely novel mathematical insight.

The company announced the finding on Monday, saying the model did not prove the conjecture outright but did push the known boundary for where the hypothesis has been checked. The work was carried out through a long, highly coordinated AI workflow that involved dozens of sub-agents and millions of tokens, and it was later reviewed by Anthropic mathematicians and verified with the Lean proof assistant.

The claim lands at a moment when AI-assisted mathematics is becoming one of the most closely watched frontiers in artificial intelligence research. A wave of recent results from Anthropic, OpenAI, and others has raised hopes that large language models may become useful collaborators in high-level mathematical discovery — while also intensifying concerns about authorship, attribution, and the reliability of machine-generated proofs.

What Anthropic says happened

Anthropic says the model was prompted by a staff member with limited mathematical training to take a serious run at the Riemann hypothesis, after which the system continued working autonomously over roughly a day and a half. During that period, it explored hundreds of potential approaches and used a multi-agent setup to sift through ideas, test them, and discard dead ends.

The company says the model evaluated 650 different ideas in total. Those attempts were distributed across 60 sub-agents that were tasked with brainstorming, checking, validating, and drafting parts of the work.

According to a footnote in Anthropic’s paper, only a small subset of those sub-agents produced the most important mathematical advances. Most were used to explore paths that turned out to be unhelpful, while others served as internal critics to test whether candidate arguments actually held together.

Anthropic’s write-up says two of the 60 sub-agents generated the key mathematical ideas, while others contributed supporting thoughts, validation, or drafting help.

Importantly, the company is not claiming a full proof of the Riemann hypothesis. Instead, it says the model improved the lower bound for the range of cases in which the conjecture has been verified. In practical terms, that means the model advanced a known line of mathematical evidence without resolving the overarching problem.

Why the Riemann hypothesis matters

The Riemann hypothesis is one of the central questions in modern mathematics because it is deeply connected to the way prime numbers are distributed. Prime numbers appear irregular, but mathematicians believe there is an underlying structure governing where they occur. The hypothesis offers a precise statement about that structure, and a proof would have major implications across number theory and related areas.

First posed in the 19th century, the conjecture has resisted all attempts at a complete proof. It is so significant that it is one of the Clay Mathematics Institute’s Millennium Prize Problems, with a $1 million reward offered for a correct solution.

That prize has remained unclaimed for years, underscoring the challenge. Even small advances are therefore notable, particularly if they come from a system that was not explicitly designed to be a theorem prover in the traditional sense.

How the problem is usually approached

Mathematicians typically attack the hypothesis using a mix of analytic number theory, complex analysis, and increasingly powerful computational checks. Because the problem is so difficult, progress often comes in the form of incremental bounds, partial theorems, or equivalences that narrow the range of uncertainty.

Anthropic’s result belongs to that category of incremental progress. It does not settle the hypothesis, but it demonstrates that an AI model can search the space of possible arguments and, with the right scaffolding, contribute something that humans consider meaningful.

How did the model do it?

Anthropic says the breakthrough came from combining a capable language model with a deliberate orchestration strategy. Rather than relying on a single stream of reasoning, the company split the work among many sub-agents that could propose ideas, evaluate them, and refine promising directions.

This approach is increasingly common in frontier AI research because it can mimic something closer to a research group than a single chatbot. One model may propose a conjecture, another may attempt to disprove it, and others may compare notes or clean up the final presentation.

In this case, the company says the model used 31 million tokens overall, a sign of the scale of the search process. Tokens are the chunks of text processed by large language models, so the figure suggests a substantial amount of computational and reasoning activity was involved.

What the sub-agents were responsible for

The paper’s breakdown offers a rare look at how these systems are being deployed in research settings. Some sub-agents were idea generators, some functioned as validators, and a small number helped write the initial paper that described the result.

• Idea generation: several agents proposed mathematical approaches.
• Key discovery: two agents produced the central advances.
• Validation: others checked whether the arguments were internally consistent.
• Documentation: a few agents helped draft the write-up.

That distribution matters because it suggests the system was not simply generating fluent text around a preconceived answer. Instead, it was participating in a dynamic loop of exploration and verification that resembles the early stages of serious mathematical research.

Why this result is attracting attention now

This latest announcement arrives amid a growing list of AI-assisted mathematical achievements. Over the past year, researchers have reported models solving or advancing problems associated with the Erdős problem set, while OpenAI has highlighted ten major results associated with its internal Astra model. Anthropic itself has also been tied to a separate advance involving the Jacobian conjecture.

That cluster of results is changing the conversation. Until recently, most public attention around large language models focused on writing, coding, customer service, or general productivity. Now, the frontier is shifting toward formal reasoning tasks that were once thought to be well beyond the reach of probabilistic text systems.

The implications go beyond mathematics. If models can help identify new structure in a notoriously difficult domain, advocates argue, they may become useful for scientific discovery in physics, chemistry, materials research, and other fields where deep reasoning matters.

What this does and does not prove about AI

The finding does not mean AI has become a replacement for mathematicians. It does, however, suggest that current models may be better at discovery than many skeptics expected, especially when paired with careful prompting, multi-agent search, and formal verification tools.

At the same time, the result does not eliminate the usual concerns around AI-generated work. Models can still hallucinate, overstate confidence, or produce arguments that appear plausible but fail under scrutiny. That is why the verification step with Lean and human mathematicians is central to the claim.

1. AI can search large idea spaces quickly.
2. Human experts are still needed to interpret and verify outputs.
3. Formal tools remain essential for proving correctness.

How do mathematicians feel about AI-generated proofs?

Mathematicians are divided, with some welcoming AI as a powerful new research assistant and others warning that it could alter the norms of the field in uncomfortable ways. The split is not just about technical quality; it is also about the culture of mathematics itself, including who gets credit for discovery.

That concern was underscored in a public declaration released in June and signed by prominent mathematicians. The statement argued that AI could weaken an important principle in the field: that a genuine proof should be tied to identifiable authors who are accountable for its correctness.

For many mathematicians, this issue is not abstract. Credit determines careers, reputations, and research trajectories. If AI systems start producing results that are accepted by the community, the profession will need new norms for authorship and responsibility.

The June declaration warned that mathematics depends on proofs being attributable to specific people who can stand behind them.

Timothy Gowers offers a different view

Not everyone sees the change as a threat. In a response to the declaration, Fields Medal winner Timothy Gowers suggested that the relationship between mathematics and authorship may evolve in ways that are less alarming than critics fear.

Gowers argued that mathematical knowledge might eventually become less tightly linked to individual names, much as many stars in the sky are unnamed. His point was not that attribution no longer matters, but that the field may adapt rather than collapse under AI assistance.

That perspective captures the uncertainty now surrounding the use of language models in theorem discovery. The same tool that unsettles traditional norms may also help mathematicians explore ideas that would otherwise remain out of reach.

What the Anthropic announcement signals for AI research

Anthropic’s result is significant because it suggests a new phase in model capability: not just answering questions, but sustaining a long sequence of interdependent reasoning steps in a formal domain. That is a much harder challenge than generating a plausible explanation or a quick solution to a textbook problem.

The company’s workflow also highlights the growing importance of agentic systems. Instead of one monolithic model performing all tasks, the system used a distributed setup in which specialized sub-agents played different roles. That design may become a template for other research applications if it proves reproducible and reliable.

Still, many questions remain. For example, how much of the progress came from the underlying model itself versus the orchestration architecture? How often would a similar setup reproduce comparable gains? And how far can these systems go before they run into the limits of current training and reasoning methods?

Key open questions

Researchers and observers will likely focus on several unresolved issues:

• Whether the progress can be replicated on other deep mathematical problems.
• Whether the result depends on unusually careful prompting or operator skill.
• Whether future models can produce not just partial progress, but full proofs.
• How formal verification will be integrated into AI-assisted mathematics at scale.

Those questions matter because the gap between “interesting” and “transformative” is still large. A model that can improve a lower bound is impressive; a model that can consistently generate publishable and correct proofs would be far more consequential.

What happens next?

For now, Anthropic’s result should be understood as an important proof of concept rather than a final answer. The company has shown that an unreleased model can contribute to a problem long viewed as beyond the reach of ordinary AI systems, but the larger scientific community will want to see whether the technique generalizes and withstands scrutiny.

If it does, the implications could be broad. Universities, labs, and AI companies may increasingly treat models as active collaborators in mathematical and scientific work, not just tools for summarizing papers or checking calculations.

If it does not, the result will still be useful as evidence that the field is moving fast. Even partial advances on a problem as iconic as the Riemann hypothesis are enough to reshape expectations about what AI can and cannot do.

Either way, the announcement adds weight to a growing argument: the most advanced AI systems are starting to participate in the discovery process itself. For mathematics, that may mark the beginning of a long transition — one that will force researchers to rethink not only what counts as progress, but who, or what, gets to make it.

Timeline of the reported advance