Skip to content

AI

Anthropic's Secret Model Spent 31 Million Tokens on Riemann

An unreleased Anthropic model reportedly pushed the Riemann hypothesis further than expected, after a non-mathematician told it to take a real stab. The $1 million bounty is still unclaimed, so no, nobody is cashing in yet.

Cover photo: Sérgio Valle Duarte .mw-parser-output .messagebox{margin:4px 0;width:auto;border-collapse:collapse;border:2px solid var( / CC BY 3.0

TL;DR

  1. Anthropic says an unreleased model significantly raised the lower bound where the Riemann hypothesis holds.
  2. A staffer without serious math training prompted it; the model then ran for a day and a half.
  3. 650 ideas, 60 subagents, 31 million output tokens. The $1 million bounty remains unclaimed.

What happened

On Monday, Anthropic announced that an as-yet-unreleased model made significant progress on the Riemann hypothesis. The model significantly raised the lower bound of solutions for which the hypothesis is known to hold true. It did not prove the thing.

The hypothesis has been open for more than 150 years and concerns how prime numbers are distributed. A $1 million bounty for a working general proof sits unclaimed. The model did not collect it either.

The setup is the funny part. According to TechCrunch's report, an Anthropic staff member without significant mathematical training asked the model to "take a real stab" at the problem. Then they walked away.

The model coordinated the work by itself over the following day and a half. In total it tested 650 different ideas, spread across 60 subagents, and burned through 31 million output tokens.

Consider the division of labor: one human typed a casual prompt, and 60 subagents did the sweating. The human's contribution was roughly "go on, then".

Why it matters

Whether AI can discover new mathematics, rather than just reproduce old textbooks, has been a long-running argument. TechCrunch says this result is likely to reopen it. Models still cannot solve the Riemann hypothesis, but they can get further than many expected.

Note what kind of progress this is. A higher lower bound is a concrete, checkable result, not a vibe. That is what separates it from the usual "AI will transform science" press release.

Caveats, because we are not animals. The model is unreleased, so nobody outside can poke at it. The claim comes from Anthropic's announcement, and our sourcing is a single report. Wait for mathematicians to chew on the details before anyone prints t-shirts.

The most interesting detail is not the model. It is that the person steering it did not need a math degree.

What it means for Europe

Not much directly: this is a US lab and a US announcement, with no specific regulatory hook or European company in the story. But frontier capability does not respect borders. European research groups and universities will be asked the same question as everyone else: what do we do when a model can run hundreds of mathematical experiments over a weekend?

What to do with it

  • Watch for verification.Look for independent mathematicians confirming or challenging the new lower bound before you repeat the claim at dinner.
  • Study the workflow.One loose prompt, one coordinating model, many subagents: that pattern is worth understanding for your own long-running tasks.
  • Mind the label."Unreleased" means you cannot reproduce this today, so treat it as a signal, not a tool.
  • Keep your own domain sharp.If a non-expert can steer a model into deep territory, your expertise is now the part that checks the answer.

Ciao, sapiens. The primes are still unsolved, so you have a few more years of job security.