$400K of GPT Tokens Got to 84%. Opus 5.5 Started Over and Finished for $24K.
Theo Browne of Ping Labs posted ts-rust on Wednesday, a Rust port of Microsoft's native TypeScript compiler, checker and language server, written entirely by LLMs. Hacker News gave it 108 points, the repo passed 680 stars in a day, and the README opens with "I've never read a line of this code." It installs as tsc-rs, takes the same flags as tsc, and claims 100% compatibility on every real-world project the team has tested, with Linux x64 and macOS arm64 builds and the Effect language-service diagnostics ported in. It is pinned to one upstream revision of microsoft/TypeScript and marked early release.
The cost story is the reason to read it. Browne first spent over $400,000 in API-priced tokens on GPT-5.6 Sol and GPT-6 Astra, running /goal loops for months. They wrote more than 1.3 million lines of Rust and never got past about 84% compatibility. Then, noticing how little Claude Code was drawing against subscription limits, Browne pointed Opus 5.5 at the same task. It had a working v0 in 10 hours. Browne assumed it had reused the Codex-written code and found it had started from scratch, getting further than Astra in a tenth of the time. Two more weeks took total spend to about $24,000 in API terms, run through Claude subscriptions at somewhere between 925% and 983% of a $200 plan's weekly limit.
Strip the fandom out and there are two facts. One is that a compiler port of this size is now a thing one person does with an agent over two weeks, for less than a junior engineer's quarter. The other is that the model mattered more than the budget by a factor of roughly 17 in cost and 10 in time on the same task with the same person driving, which is the cleanest single-task harness-versus-model comparison anyone has published, and it points at the model. Make of the "/goal loops for months" detail what you will; a loop that runs for months without converging is a harness problem as much as a model one.
The README's own warning applies: an early release, unknown correctness in the tail, and a maintainer who has not read it. That is also, increasingly, what shipped software looks like.
Links: github.com/pingdotgg/ts-rust, news.ycombinator.com/item?id=50000676
← Back to all articles
The cost story is the reason to read it. Browne first spent over $400,000 in API-priced tokens on GPT-5.6 Sol and GPT-6 Astra, running /goal loops for months. They wrote more than 1.3 million lines of Rust and never got past about 84% compatibility. Then, noticing how little Claude Code was drawing against subscription limits, Browne pointed Opus 5.5 at the same task. It had a working v0 in 10 hours. Browne assumed it had reused the Codex-written code and found it had started from scratch, getting further than Astra in a tenth of the time. Two more weeks took total spend to about $24,000 in API terms, run through Claude subscriptions at somewhere between 925% and 983% of a $200 plan's weekly limit.
Strip the fandom out and there are two facts. One is that a compiler port of this size is now a thing one person does with an agent over two weeks, for less than a junior engineer's quarter. The other is that the model mattered more than the budget by a factor of roughly 17 in cost and 10 in time on the same task with the same person driving, which is the cleanest single-task harness-versus-model comparison anyone has published, and it points at the model. Make of the "/goal loops for months" detail what you will; a loop that runs for months without converging is a harness problem as much as a model one.
The README's own warning applies: an early release, unknown correctness in the tail, and a maintainer who has not read it. That is also, increasingly, what shipped software looks like.
Links: github.com/pingdotgg/ts-rust, news.ycombinator.com/item?id=50000676
Comments