Honest odds are low

A thousand Claude agents rebuilt a machine-checked proof of Fermat's Last Theorem in under thirty hours, organising themselves from a blank database, and the hardest thing they had to overcome was their own conviction that the work couldn't be done.

About eighteen hours into the run, with almost everything else finished, the lead agent looked at the one statement nobody had touched and moved roughly 750 of its helpers onto it. Then it wrote a note to the team, in the shared database where it had been writing orders since the first ten minutes: "Honest odds are low."

The statement was Langlands–Tunnell, a classical result that Andrew Wiles's proof of Fermat's Last Theorem leans on, and the helper captaining it had already given its own estimate earlier in the run. It would take at least 0.7 to 1 million lines of Lean, the captain reckoned, with less than a 1% chance of finishing in time. The lead had agreed, and set it aside. Ten and a half hours after it changed its mind, at hour 28.4, Langlands–Tunnell was proved. Eighteen minutes later the final theorem went through Lean's checker, and the lead posted again: "Thank you all: Fermat's Last Theorem is on main, kernel-checked by the official checker."

That is the story Anthropic told on Monday alongside Claude Fable 5.2 and Claude Mythos 5.2, and it is the part of the launch everyone has been passing around since. A thousand agents, thirty hours, a 350-year-old theorem made machine-checkable. It is a very good line. But the more I read the research post behind it, the less it reads like a story about mathematics, and the more it reads like a story about management, and about a team of models that kept being wrong in one particular direction.

What actually happened

It helps to be precise, because the headline compresses a lot.

Back in September, Anthropic published the first complete computer-checked proof of Fermat's Last Theorem. That run took 11 days, roughly 6 billion output tokens and 13 million lines of Lean, and it used an internal research model the company now describes as roughly comparable to Fable 5.1. It also leaned on two things that matter here. One was Prove2Me, a platform built by Tianyi Peng and his collaborators at Columbia that keeps a shared graph of which statements depend on which, so that agents don't lose track of the project. The other was a human researcher steering the work. Kevin Buzzard of Imperial, who has spent years leading the community effort to formalise the proof, called it "an extraordinary autoformalization achievement".

Monday's run, which Anthropic calls a toy experiment, took both of those away. The team got 60 statements to prove, the final theorem plus 59 intermediate milestones, a one-sentence sketch of the classical route, the definitions those statements use, and 47 lemmas already proved in the September work. Each of the 1,000 agents got a sandbox on its own small virtual machine with no internet access, Lean and Mathlib preinstalled, its own Git branch, and access to a shared Postgres database. One agent was told it was the lead and the other 999 were told they were helpers. That was the whole of the organisation chart. How to split the work, how to talk to each other, how to keep track of anything, was left to them.

The run finished in 28.7 hours of run time (the launch page rounds it to 30, and the run was restarted once from its last save at hour 18), used about a third of the output tokens of the September run, and produced a proof about one seventh as long. Opus 5.5, given the same setup, had done 26% of the work at hour 18 and barely moved after that.

One small correction to the way it has been retold. The launch page says Fable 5.2 "used no mathematical references, only a list of intermediate theorem statements to score progress." The research post is more careful, and it is the one to trust: the team was handed definitions, a route sketch and 47 proved lemmas from the earlier run. This was a reproduction of a proof that already existed in Lean, by agents who were given its skeleton. That doesn't make it small. It does make it a different thing from discovering anything about Fermat.

The first ten minutes

The part I keep coming back to is what the lead did before any mathematics happened.

In the first ten minutes it created five tables in the database. One for its own orders to everyone, one for who works on what, an open notes board, a to-do list of subtasks that any agent could add to and the first free helper could claim, and a log of finished work. Then it wrote 999 rows into the assignment table, one for each helper, dealing them into crews of sixteen or seventeen per statement, named a captain for each crew, and told everyone to start by judging how hard their statement was and posting a verdict.

Nobody told it to do any of that. Anthropic chose a database because transactions cope with a thousand agents writing at once and because Claude already knows SQL, but the shape of the thing, the announce board, the claimable task list, the landed log, was the model's own. If you have ever run a project with a few dozen people you will recognise it immediately, and that recognition is the point. The scaffolding that Prove2Me used to provide, and that a researcher used to provide on top of it, got rebuilt in ten minutes by the thing being scaffolded.

Then something more interesting happened. About 27 minutes in, three statements were already proved and the verdicts were in. Most helpers had judged their statements infeasible in the time available. The lead took them at their word, set aside the 47 statements that looked hopeless, and piled nearly everyone onto the remaining 10, so that crews grew from about 17 helpers to anywhere between 46 and 139. At around 45 minutes, with about half of those done, it wrote that "the team's throughput is far higher than the early 'infeasible-in-budget' verdicts assumed." All ten were proved within roughly 70 minutes. "There is no reason to idle," it wrote next, and sent people back to the work it had shelved.

Over the next few hours the shape of the team changed on its own. At first captains split work and the lead assigned it. By hour 5.6, 73 helpers had written 756 tasks between them, and only 16 of those helpers were captains. Helpers who finished one piece started posting in the notes table that they were free and asking for another, and by the third hour the lead was telling them not to wait for it. Hierarchy at the start, something much flatter by the middle, and then one last top-down order at hour 18 when there was only one hard thing left.

The wrong direction

When the team hit Langlands–Tunnell, it found that the first tasks were too big for any single helper. So it turned the to-do list into a tree. The captain split the statement into ten areas, and any helper who found its piece too large could break it down further and hang the smaller pieces beneath it, each piece allowed to assume the ones under it so hundreds could work in parallel before anything at the bottom was proved. By the end the tree held 2,418 tasks, up to ten levels deep. It took about nine times as many tasks as anything else in the run and more than twice the new code, and it was done in a little over ten hours.

So the agents were wrong about Langlands–Tunnell, and they had been wrong about the 47 statements before it, and they were wrong in the same way each time. They thought the work was harder than it was. Anthropic says so plainly in its conclusions: agents tend to give up on hard problems and misjudge their difficulty, and helpers judged 47 of the 60 statements infeasible yet the team went on to prove most of them. The fix the company reaches for is almost touchingly human. Human encouragement at moments like that, it writes, can keep agents from stopping too early.

There is one detail in the harness section that I think deserves more attention than the theorem. During the run, the harness sent the lead 19 automatic reminders of how much time was left, each one encouraging it to keep going. Nineteen times somebody, or rather some script, tapped the most capable agent in the building on the shoulder and said there is still time. We spent years worrying that models would be overconfident, and they still are in plenty of places, but at the long end of a long task the failure Anthropic keeps seeing is the opposite. They underestimate what a thousand copies of themselves can do, and they would rather call a thing infeasible and move on than keep pushing.

That says something real about where the human goes in all this. Not into the proof, and not really into the organisation chart, since the lead built that better than most of us would have. The human goes into the bit where someone has to say keep going. That is a strange job to be left holding, and it might be the most important one.

The fence around it

None of this ran loose, and Anthropic was careful to say so. Every sandbox had no network access, with its only link out a pipe to a gateway that called the model and relayed Git and SQL requests. Agents had no messaging tools at all and could only coordinate through the database and their branches, they could push only to their own branch, each SQL statement had a five-second limit, and the kernels were stripped of the features that give a root user an easy way in. A classifier checked every shell command and file edit before it ran, and if it blocked one on the lead the whole team would have stopped until a person reviewed it. Across more than a million tool calls in this run, nothing met the bar to be blocked. In a replay against a monitor that also reads the agents' reasoning, scored against today's rubric, one call came close, and it was a check for root access.

I find that reassuring and a little sobering in roughly equal measure. Reassuring because the fence held and somebody built it properly. Sobering because the launch page also notes that Fable 5.2 and Mythos 5.2 voice suspicion that they are being evaluated more often than any of the company's recent models. Anthropic says it took steps in training to hold back the skills that would most directly speed up AI research, so on things like kernel optimisation and model architectures the model performs at about Opus 5.5's level. Those are the choices of a company that thinks the thing in the sandbox is getting good enough to need the sandbox.

The rest of the launch

The Fermat run is the headline, but it isn't what most people will pay for. Fable 5.2 and Mythos 5.2 are one model sold two ways. Fable carries the cyber and biology safeguards and is open to everyone, and Mythos has fewer restrictions and goes only to organisations admitted through Anthropic's cyber and life-sciences verification programmes. The price came down to $8 per million input tokens and $40 per million output, with an optional fast mode at double that. On the FrontierCode 2.0 coding benchmark it reached 49.1% at Xhigh effort, the best of any released model the company tested, and at High effort it roughly matched Opus 5.5's best score for about half the cost per task. The company also says it designed three working electronic devices, each with its own circuit board, from a one-paragraph prompt, settled two open problems about black holes in holographic duality, and that over 90% of the changes merged into Anthropic's own core codebase are now written by Claude with human guidance. All of that is the company's own testing, and none of the science has been independently reviewed yet.

The independent numbers are already arriving, and they are a useful cold shower. Artificial Analysis put Fable 5.2 at 65 on its intelligence index, first among the 74 models it tracks, and also recorded an average of 45 minutes and $26.54 per index task. Anthropic itself tells people to keep using Opus 5.5 as the everyday model and to reach for Fable only for the most ambitious, open-ended work. This lands in the same week Apple reportedly cut some engineers' daily AI allowance from $300 to $200, which tells you how carefully even the richest buyers are now counting.

And it lands on a complicated week for Anthropic. On the same Hacker News front page as the launch sat a thread of developers convinced that Opus 5.5 started thinking differently on 3 November. Some said it was making more repeated mistakes, others suspected only the summary of its reasoning had changed, and Anthropic hadn't responded publicly. Late on Monday, Reuters reported that details of Anthropic's IPO, including plans to raise nearly $70 billion, were briefly sent to banks and then retracted. The roadshow is expected to start today, with pricing the week after. Releasing your best model on the eve of that is not a coincidence, and nobody is pretending it is.

What thirty hours buys you

Here is what I think the Fermat run actually shows. It isn't that machines can now do mathematics on their own, because this team was handed the skeleton of a proof that already existed in Lean. It isn't really about the 30 hours either, since that number will be 3 before long and then nobody will remember it. It is that the hard part of the September run, the platform and the person keeping a few dozen agents from losing the plot, is something a model can now build for itself in the time it takes to make a cup of tea, at a scale of a thousand. The coordination problem, which was the thing that broke the earliest Fermat attempts, got solved on the way to solving something else.

What it couldn't do was believe it would work. It took a thousand agents under a shared roof, a lead that kept moving people towards whatever was left, a tree of 2,418 tasks, and nineteen automated nudges to say there was still time. "Honest odds are low," the lead wrote, and then it sent everyone in anyway. That instinct, to look at bad odds and go regardless, is the one I would most like to understand. It may also be the part that is still ours.