Skip to content
Week 47News, and what it means for you

Science

Claude solved two open problems about black holes

Anthropic’s AI agents worked on open problems in physics for six weeks on their own and cracked two. How that work was made trustworthy is at least as interesting as the result.

Anthropic reports that its AI model Claude has solved two open problems in theoretical physics, both about black holes. Agents worked on them on their own for six weeks, and a physicist checked every main result.

It started in August as a side project of Anthropic researcher Hongbin Chen. He asked Claude for open problems about black holes and the bootstrap, a method for testing theories, and got 22 candidates within a few hours. He set a separate agent to work on each problem: an AI that calculates, writes code and reports back on its own. The agents could read each other’s progress and work together on related problems. They ran in the background for about six weeks with no special setup, and almost every step was done twice by agents working independently.

The first result completes a calculation that was set up in 2003 but never finished. It was known that in a large theoretical universe, hot gas turns into a black hole abruptly, the way water starts to boil. Claude worked out that the same holds in a very small universe. The second result gives the exact ceiling on the mass of the lightest black hole in a simplified universe with three dimensions, a question physicists had been chipping away at since 2009. That proof was also checked in Lean, software that verifies every step of a mathematical proof.

Anthropic announced the results on 9 November with the launch of Claude Fable 5.2 and published the full write-up on 17 November. The two papers are preprints on hexagonmath.org, a server for mathematics and physics. They have not yet been peer reviewed, and we found no independent physicists who have publicly rechecked them yet. Anthropic itself adds that neither result directly describes our own universe or has a practical use.

The agents ran in the background, with no special setup, for about six weeks.

Anthropic, 17 November 2026

When hot gas becomes a black hole

In 1983 Stephen Hawking and Don Page showed that in anti-de Sitter space, a theoretical universe that behaves like a box, a black hole only becomes the preferred state above a certain temperature. The switch is abrupt. According to holography, such a universe has the same physics as a particle theory without gravity, here N=4 super Yang-Mills, so that theory should have a matching transition.

In 2003 Ofer Aharony and colleagues showed that in a very small universe the answer hangs on a single number, which only appears deep into the calculation, buried in thousands of diagrams. In 2005 the same group worked it out for a simpler related model. The N=4 version is more than ten times as large and was left undone. Claude found a negative number, which means an abrupt transition, just as Hawking and Page found.

A ceiling for the lightest black hole

The bootstrap rules out theories without fully solving them: you only demand that they obey a few hard rules of consistency. In 2009 Simeon Hellerman applied it to gravity in three dimensions and proved a ceiling of c/6, where c is a fixed number of the theory. In 2019 Tom Hartman and two colleagues reached c/8.5, through a surprising link with packing spheres as tightly as possible. Numerical work pointed to about c/9.1.

Claude found an exact formula, c divided by about 9.1266, and proved that this version of the method cannot go further. So it cannot bring the ceiling down to c/12, the mass at which black holes first appear. The big questions about quantum gravity remain open, Anthropic says.

How the agents got there

An earlier agent had failed on the ceiling in 61 parallel attempts. The breakthrough came from teamwork: one agent spotted a pattern in the numerical results that physicists had missed, and an agent working on sphere packing turned it into a method. Two other agents later reached the same ceiling independently. According to the preprint, an internal Claude research model found it first, and the publicly available Claude Opus 5.5 later reproduced it on its own.

For the first result, a second confirmation came from an unexpected corner. Weeks after the agent had finished, a physicist who had separately been using Claude on the same problem shared his result. It matched, but was reached by a different route.

Our analysis

The real news is how to make agents work reliably for weeks

An AI that calculates on its own for six weeks makes mistakes along the way, just like a person. Anthropic’s setup catches them with three layers. Independent double work: two agents that calculate the same thing separately rarely make exactly the same mistake. A formal check: Lean only accepts a proof in which every step holds. And an expert who rechecks every main result. Together these layers turn a fast, tireless calculator into a trustworthy one.

What the coverage underplays: most of the work was not a flash of genius but persistence. The first problem sat for over twenty years because it was too laborious, not because nobody knew how to do it. That kind of work, large, precise and endless, suits agents well. And these are two of the 22 candidates: not every open problem is within reach of agents either. Until independent physicists have gone through the preprints, it remains a strong claim by the maker itself.

Our expectation: the same approach will make the difference in organisations for work that takes weeks, such as cleaning up a large dataset, documenting an old system or preparing a migration.

What stands out

  • Help from another field The breakthrough on the ceiling came from an agent working on packing spheres, not on black holes.
  • Failure was part of it An earlier agent failed in 61 parallel attempts. Only teamwork between agents produced the answer.
  • Not every result was checked the same way The ceiling was checked in Lean. For the transition to a black hole, Anthropic cites calculating everything twice and comparison with known cases.
  • Science for its own sake Anthropic itself says neither result directly describes our universe or has a use. The news is the way of working.

What it means for you

AI agents can keep working on hard problems on their own for weeks, if you set them up so that every result can be checked.

  • Have important work done twice Two agents working independently catch each other’s mistakes. If their results differ, you know where to look.
  • Let software check what software can check Tests, calculation rules and comparisons with known results run automatically, every time.
  • An expert approves the end result Have a colleague who knows the field recheck every important result before it is used.

Where we come in · Solutions

We develop agents whose work you can check

Flaire develops tailored AI agents for work that takes a lot of patience, such as pulling invoice details from PDFs or entering orders into your administration or ERP. We set up the checks from the start: double work where it matters, automatic checks on the results and a colleague who approves the final step.

What work would you hand to an AI agent for weeks on end?

Tell us which work keeps getting shelved because it is too laborious. We’ll think it through with you and always reply within one working day.

More from week 47

All news

Search our services, the news and the rest of the site.