Theo ships tsc-rs after $400k of Codex got nowhere
Five months of agent work, two models, and a 20x gap in what the winning run cost.
Theo released tsc-rs, a complete rewrite of the TypeScript compiler, type checker and LSP in Rust, built entirely by agents over five months. It is open source and ships as a drop-in replacement for tsc.
I burned ~$400k of Codex tokens and got nowhere. Burned ~$20k of Opus and got there in 2 weeks.
I have not read a single line of the code.
The number everyone is staring at is the split. Theo says he burned roughly $400,000 of Codex tokens and got nowhere, then burned roughly $20,000 of Opus and got there in two weeks. Same project, same person steering, two orders of magnitude apart on the run that actually finished.
He also says he has not read a single line of the code.
What the $20k actually was
In a follow-up, Theo put the Opus spend in plan terms, calling it the equivalent of about ten weeks of usage on the $200 Claude subscription. That squares with his earlier claim that the $200 tier buys far more inference than it costs. On that math the finished compiler is a few months of a consumer subscription, and the $400,000 is the price of having picked the wrong model for long enough to notice.
The honest caveats
Theo put them in the announcement himself. Bun shipped `bun check` on the same day, Jarred shared numbers with him early, and Theo calls it incredibly impressive and probably the right choice for most apps using Bun. tsc-rs is only faster if you are running the Effect ts checks. He also notes `bun check` is missing incremental checks, which is the one gap he thinks leaves tsc-rs as the best drop-in option right now.
His pitch for tsc-rs rests on compatibility rather than raw speed. Drop-in for tsc, 100% compatibility, fully wasm-ready, Effect ts mods built in. His closing instruction to readers was to try it and tell him how broken it is, which is the correct posture for something nobody has read.
The business read is about model selection, not about Rust. A solo founder running agents on a hard, well-specified problem now has a cost variable that moves by 20x depending on which model is doing the work, and five months of sunk tokens is how long it took one very experienced user to find that out. Nobody has reproduced the comparison, and it is one project from one builder.
This was the equivalent of ~10 weeks of usage on the $200 Claude sub btw