Anthropic rewrote its Messages API in Rust with no Rust devs

@somi_ai says Anthropic rewrote its Messages API in Rust with zero Rust developers on the team. The claim is sourced to @katelyn_lesse on @latentspacepod. Claude wrote the code. Humans wrote the plan and checked that old and new matched on real traffic. CPU dropped 40 to 50%.

swyx, who runs the show, confirmed it in his own words, saying Anthropic really vibecoded their entire backend to Rust and none of you noticed. He added that people were trying to quote his own show notes back at him.

swyx

@swyx

X

its true, Anthropic really vibecoded their entire backend to Rust and none of you noticed

What is actually interesting here

Not the language choice, and not the speed. The thing worth copying is the division of labour. The model produced the code. The humans produced the plan and the verification, running old against new on real traffic to confirm behaviour matched before anything mattered. That is the part nobody can delegate yet, and in this telling nobody tried to.

The conventional wisdom @somi_ai is swinging at is the old rule that you never do the big rewrite. That rule exists because rewrites burn months of engineer time and then ship subtle behaviour changes that nobody catches until customers do. If the typing is close to free, the cost structure of that rule changes. The verification cost does not change at all. It arguably goes up, because you now have more generated code to check than you would have written by hand.

The caveats worth holding

This is a secondhand report of one rewrite at one company, relayed on a podcast and amplified on X. The 40 to 50% CPU figure is Anthropic's own number as reported, not something anyone outside has measured. And the company in question makes the model that did the work, which is both the most interesting thing about the story and the reason to read the number with one eyebrow up.

There is also a scale point hiding in it. A Messages API is a high traffic surface with a very clear correctness test, which is whether responses match. That is close to an ideal shape for an agent rewrite. A system with fuzzy outputs, no traffic to replay against and no obvious diff is a much harder sell, and nothing in the sources speaks to that case.

For a solo founder the read is narrower than the headline suggests. You probably do not have a Rust rewrite waiting. You might have an old service you have avoided touching because the typing looked like three weeks. The question to ask is not whether the model can write it. It is whether you have a way to prove the new one behaves like the old one.

Somi

@somi_ai

X

Humans wrote the plan and checked old vs new matched on real traffic.