Fermat’s Last Theorem was written in the margin of a book in 1637 and proved by Wiles in 1994—a span of more than 350 years. Wiles himself spent seven years working in seclusion. Early this morning, Anthropic announced Claude Fable 5.2, saying that 1,000 agents spent 30 hours turning the entire proof into Lean code that a machine can check line by line, with almost no human intervention. The same 1,000 agents using Opus 5.5 spent 18 hours reaching 26%, then barely made any further progress.

Miners know all about throwing more machines at a problem. But that Opus group shows that if a single machine isn’t powerful enough, even a thousand machines can get stuck.
The benchmark scores have been circulating everywhere: Artificial Analysis Intelligence Index—Fable 5.2 Max: 64.6, Opus 5.5: 59.1, GPT-6 Astra Max: 56.9. I’d like to offer a few thoughts from the crypto world.
Let’s start with verification. The proof is still Wiles’s; what’s new is a version that a machine can run. Whether it’s correct is something you can find out by running it, without having to trust the reviewers. That’s “Don’t trust, verify.” Formal verification of smart contracts has been around for a while, but it’s expensive and slow, so only top-tier protocols can afford it. A proof on the scale of Fermat’s Last Theorem can be completed in 30 hours. I suspect that in the future, formal verification of a lending protocol will mainly be a matter of paying the compute bill.
Now, the audit story. This time, it’s one model with two names: Fable is available to everyone, but the official announcement says most security-related requests will be blocked. The version that isn’t blocked is called Mythos, and it’s only available to vetted organizations. The announcement cites a security company’s claim that, among the vulnerabilities confirmed by Mythos 5.2, one in six hadn’t been found by any other frontier model. The footnote says this figure comes from a single testing firm; Anthropic’s own tests found that its ability to find vulnerabilities was about on par with Opus 5.5. For DeFi, the next thing to watch is which audit firms have access to Mythos.
Next, pricing. The API costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 in the previous generation—a 20% price cut. Output can also be up to 30% faster. But according to The Decoder, running the AA benchmark at the highest tier costs $26.50 per problem, compared with $12.99 for Fable 5.1 and $10 for Opus 5.5. It’s a bit like gas prices going down while a transaction burns twice as much gas. According to Fortune, Anthropic itself recommends using Opus 5.5 for everyday tasks.
One last small item—friends in the TapeOut community might want to take note. In an official demo, given a prompt consisting of a single sentence, Fable 5.2 designed three usable electronic devices: a voice assistant, a USB conference microphone, and an electronic pet. Each included a four-layer circuit board, casing, and software. The boards were ready to send to production in under a day, and the finished devices needed no circuit changes. These are circuit boards, still a long way from tape-out, but “one sentence to one board” is already a reality.
Fortune says this could be Anthropic’s last major release before going public. Which model do you mainly use right now, and would you switch for this one?
—
This article was first published on X @musinoot