A 40% CPU Win in a Claude-Assisted Rust Port: What Actually Moved
An Anthropic team with no prior Rust experience reportedly used Claude to refactor a core API, and CPU dropped about 40% while latency fell 20%. That is a headline and one sentence — no workload, no hardware, no first-party write-up I could find. So this is not a reproduction. It is an autopsy. The argument: a 40% CPU cut from a cross-language port is almost never a language win. It is three levers — fewer allocations per request, a concurrency model that stops oversubscribing cores, and fewer full-payload copies at the boundary — plus the fact that a rewrite forces you to do that work anyway. In the video: the code shape such a port usually produces, where every item gets its own heap string before being collected and joined; the buffer-reuse version that reserves capacity once and writes into a single buffer; and the before/after comparison across allocations, reallocations and copies. What the report leaves out: the source runtime, the request mix, the hardware, the measurement window, and which percentile the latency figure refers to. Treat crate names, versions or benchmark harnesses quoted from syndicated copies as unverified unless you can trace them to the original article or a first-party post. Full write-up: https://hackyjs.com/posts/autopsying-anthropics-40-cpu-win-which-levers-actually-moved-in-the-claude-assisted-rust-port