The team at Ping Labs had Microsoft's TypeScript compiler, which is now written in Go, ported to Rust. Claude Opus 5.5 did most of the work. The project is called tsc-rs (repository ts-rust), and its cost of about $24,047 contrasts sharply with the more than $400,000 an earlier attempt with other models had used up.

What happened

The project's README opens with the author saying he wanted to see whether language models could port the TypeScript compiler, the checker and the language server to Rust. His answer: they can. In total, the README says, it cost more than $420,000 in tokens, "but you could probably have done it for about $20k".

Two attempts

First the author used many OpenAI models. By his own account he spent over $400,000 in API-priced tokens in total, with GPT-5.6 Sol and GPT-6 Astra. The models wrote more than 1.3 million lines of Rust over several months and, in his words, never got past about 84% compatibility.

Then he tried Claude Opus 5.5, because he saw how little of his Claude Code limits were being used. After ten hours there was a working first version (v0). He assumed Opus was reusing the code of the OpenAI models; that was wrong: Opus 5.5 started from scratch and got further than Astra in a tenth of the time. He let it keep going. Total spend was about $24,047 of API cost over two weeks. Because he used Claude accounts, that works out, by his calculation, to between 925% and 983% of his weekly limits on the $200 plan.

The author also writes that he has never read a line of this code. Everything below a line he calls "The Slop Line" was written by his language models, not by him.

What the result can do

tsc-rs is a direct port of Microsoft's native TypeScript compiler, which is written in Go. It keeps the algorithms and behaviour and offers the same command line (tsc), the same language server and the same API. It is pinned to a specific state of microsoft/TypeScript (8 October 2026, TypeScript 7.1.0-dev) and was compared with Go at that state.

  • Same results: TanStack Query core and Hono produce diagnostics identical to Go's. The first report cited 181,711 passing Go tests; the current README cites 184,773.
  • Faster: On 60 open-source projects, type checking takes about half as long as in Go (geometric mean).
  • Real projects: On 120 open-source repositories the command-line output differs from Go's only in the known problems and where Go's own output varies from run to run.

Background

The author measured how fast the checking runs on 13 October 2026 on machines with 16 processor cores. For the full type check of T3 Code, once with and once without Effect diagnostics, the ranking is as follows (sum of the medians in seconds):

Checker With Effect diagnostics Without Effect diagnostics
tsc-rs with --fast 9.77 s 6.91 s
tsc-rs 14.30 s 7.95 s
tsc 7 (Go) 42.55 s 19.45 s
bun check 59.24 s 6.24 s
tsc 6 (JavaScript) 191.00 s 87.87 s

On six open-source applications, tsc-rs 0.6.0 is on average 2.10 times faster than tsc 7, and 2.90 times with --fast. For VS Code, for example, tsc 7 takes 9.57 seconds, tsc-rs 4.81 seconds and with --fast 2.64 seconds. The table also shows what the README openly states: without Effect diagnostics bun check is fastest at 6.24 seconds, and on four of the six applications Bun is ahead. The time comparisons do not mean that all tools check identically; the diagnostics differ.

The --fast switch uses one checker per processor core, at most 16. That makes checking faster but costs resources: in the measurements on 16 cores, processor time rose by up to 2.2 times and peak memory by up to 1.9 times.

The README also lists known problems. In some monorepos tsc-rs can write output for more files than tsc, and in builds with several projects the result can depend on timing. In the editor, memory grows slowly during long sessions, by 3.1 to 4.6 MiB per 1,000 edits, against 2.6 to 2.8 for tsc. There are binaries for Linux x64 and arm64 and for macOS arm64; from the first release after 0.2.0 also for macOS x64 and Windows.

More on this topic:

What this means

For developers working with TypeScript, tsc-rs is above all an experiment that can be tried out. The README explicitly calls it an early release but states full compatibility in every real-world project tested. The way to try it is npm install -D tsc-rs. Anyone testing it should know the known problems, especially for monorepos.

For the question of what AI programming can achieve, the case is instructive because it compares two runs with different models on the same goal. The figures come from the author himself: over $400,000 and several months to reach about 84% compatibility with OpenAI models, against about $24,047 and two weeks with Opus 5.5. This is not a controlled study, since these are two individual attempts with different tools and billing. But the figures show how much effort and outcome can differ in such large undertakings.

For the question of quality a tension remains. The author says he has never read the code, yet delivers tests and comparisons: he points to tests that come with the Go code and to comparisons with the Go version on many projects. Anyone who uses such code therefore relies on those checks, not on a review.

A note on interpretation: the measurements come from the project's author; the sources evaluated found no independent repetitions. The speed-up depends on the measurement: an earlier summary gave on average 1.61 times tsc 7 and 11.4 times tsc 6; the current README gives 2.10 times tsc 7 for version 0.6.0.

Fre­quently asked questions

What is tsc-rs?

A port of Microsoft's native TypeScript compiler, which is written in Go, to Rust. According to the README it keeps the algorithms and behaviour and has the same command line (tsc), the same language server and the same API.

How much did the port cost?

The README puts the part done with Claude Opus 5.5 at about $24,047 of API spend over two weeks. The earlier attempt with OpenAI models cost over $400,000 according to the README; in total it names over $420,000.

Can tsc-rs be used in pro­duc­tion?

The README calls the project an early release. It states full compatibility in every real-world project tested, but lists known problems, such as differences in some monorepos and slowly growing memory in the editor. The author writes that he has never read a line of the code.

How fast is tsc-rs?

According to the README it is on average 2.10 times as fast as tsc 7 on six open-source apps, and 2.90 times with the --fast switch. On 60 open-source projects, type checking takes about half as long as in Go. These are the author's measurements from 13 October 2026.