Theo Spent $400,000 on Codex and $24,000 on Opus to Port tsc to Rust

Theo Browne's new TypeScript compiler comes with a line most open-source READMEs would never print: "I've never read a line of this code." The project, tsc-rs (the repository is called ts-rust), is a Rust port of Microsoft's native TypeScript compiler, type checker and language server. It is published under the pingdotgg organization on GitHub, MIT-licensed, and installable from npm today. According to the README, the agents that wrote it got there twice. The first time cost more than $400,000 in API-priced tokens and failed. The second cost about $24,047 and worked.

Browne, who runs Ping Labs and the T3 developer brand, announced the project on X on October 7. "I burned ~$400k of Codex tokens and got nowhere," he wrote. "Burned ~$20k of Opus and got there in 2 weeks." The README gives the more precise figure.

Two attempts, one finish line

The target was not TypeScript in the abstract. It was Microsoft's Go rewrite of the compiler, the codebase that ships as TypeScript 7. The README describes tsc-rs as "a direct port" that keeps the Go version's algorithms and behavior and exposes the same tsc command line, language server and API. The port is pinned to a 7.1.0-dev upstream revision.

The first run used OpenAI models. The README names GPT-5.6 Sol and GPT 6 Astra, says they "wrote over 1.3m lines of Rust," and says the result "never got past like 84% compat." Browne puts that attempt at more than $400,000 in API-priced tokens.

Then he tried Anthropic's model. "I figured it'd be fun to throw Opus 5.5 at this," the README says. Claude Opus 5.5 started from scratch rather than continuing the earlier code, and had a working first version in 10 hours, according to the README. The full port took about two weeks and roughly $24,047 of API-equivalent spend.

That number needs a footnote, and the README supplies it. The Opus work ran on Claude subscriptions, and Browne reports usage at between 925% and 983% of a $200 plan's weekly limits. As developer Nikoloz Turazashvili noted in a DEV Community breakdown, both headline figures are usage valued at API list prices, not invoices. The comparison between the two attempts still holds. The dollar amounts are a unit of measure, not a bill.

What the tests actually say

The claim that will travel is "All 181,711 ported Go tests pass." That is the Go compiler's own test suite, translated to Rust along with the code. The README adds that TanStack Query core and Hono type-check with diagnostics identical to the Go compiler's, and that across 120 open-source repositories, tsc-rs output differs from Go's only in a short list of documented problems. The README also claims "100% compatibility in every real world project we have tested."

None of this has been independently reproduced yet. Turazashvili flagged that he had not run the compiler himself, and that the README pins a September 29 upstream revision while a metadata file in the repository names a June revision as the default. AICoder reported that, by Browne's account, four of the first five issues filed against the project turned out to reproduce upstream TypeScript behavior rather than porting bugs.

The more useful caveat is about the shape of the task. Porting a compiler that already exists, with a large test suite and a reference implementation to diff against, is close to the ideal problem for a coding agent. Every failure has a known right answer. Turazashvili made the same point: the agents had a strong reference to work against, which says less about how they would do on software with fuzzy requirements.

Faster than Go, slower than Bun

The README's benchmarks were run by the author on an Apple M4 Pro with 48 GB of RAM, five measured runs after one warm-up, with incremental checking off. Across six open-source apps, the geometric mean puts tsc-rs at 11.4 times faster than TypeScript 6 and 1.61 times faster than TypeScript 7. On VS Code, a full check took 4.20 seconds in tsc-rs against 6.84 seconds in TypeScript 7 and 54.56 seconds in TypeScript 6. On Browne's own T3 Code projects, tsc-rs finished in 7.25 seconds against TypeScript 7's 16.10.

The README does not hide the faster option. Bun's type checker posted a 20.9 times geometric-mean speedup over TypeScript 6 and beat tsc-rs on five of the six apps, including 1.62 seconds on VS Code. The one place tsc-rs pulls ahead is with Effect, the TypeScript library. tsc-rs builds Effect's diagnostics into the same check, and with them enabled it ran T3 Code in 11.13 seconds against 21.07 for TypeScript 7 plus the separate Effect tool. Turazashvili pointed out that the most dramatic number in circulation, roughly 12.5 times, compares tsc-rs against TypeScript 6 with an older Effect plugin.

Early release, by its own label

The README calls this "an early release," and its known-problems list is specific. Some monorepos produce spurious TS6059 errors or extra output files. Build mode can read stale or missing outputs across projects that lack references. Editor memory grows about 20 MiB per 1,000 edits. Binaries exist only for Linux x64 and macOS arm64, and "Windows is not available yet." Effect's editor features, such as quick fixes and hover, are not ported. There is also a WebAssembly build, one of the stated reasons for the project alongside speed, model testing and, per the README, "Memes."

The open question is ownership. Microsoft keeps changing the Go compiler, and a port is only as current as whoever tracks upstream. Browne's answer, implicitly, is the same agents that wrote it. That is the real experiment here. A code generator that can produce a compiler in two weeks is a striking result. A code generator that can keep one in sync with Microsoft for two years would change how infrastructure gets maintained. The first claim now has a public repository behind it. The second has not been tested yet.