ts-rust (aka tsc-rs): Rust port of the TypeScript compiler, written by LLMs
A · Strong evidence Public repository with a large, test-verified LLM-written port (181,711 ported tests passing, benchmarks on named open-source apps with commit hashes) and disclosed token costs; a measured practitioner case, though the cost narrative is the author's own account.
Summary based on the full text·Original: about 15 minutes to read
In short
ts-rust (npm package tsc-rs) is a Rust port of Microsoft's native TypeScript compiler, checker and language server, written by LLMs; its author says he has never read a line of the code. OpenAI models spent over $400,000 in API-priced tokens and wrote over 1.3 million lines without passing about 84% compatibility. Anthropic's Opus 5.5 started from scratch, had a working version in 10 hours, and reached the released state for about $24,047 of API-equivalent spend over two weeks. The port matches the Go compiler's output on tested projects and type-checks about twice as fast.
Key numbers
- $400,000+API-priced tokens spent on GPT-5.6 Sol and GPT 6 Astra, which never got past about 84% compatibility
- ~$24,047API-equivalent spend for Opus 5.5 over 2 weeks, starting from scratch
- 181,711ported Go tests that pass
- 13.4xgeometric-mean speedup of tsc-rs over tsc 6 on six open-source apps (tsc 7: 7.1x)
As reported by the source.
English·Web page·opens github.com
What it reports
- OpenAI run: GPT-5.6 Sol and GPT 6 Astra, over $400,000 in API-priced tokens, over 1.3 million lines of Rust over several months of goal loops, about 84% compatibility at best.
- Opus 5.5: working v0 in 10 hours; "got further than Astra in 1/10th the time"; about $24,047 over 2 weeks, 925%-983% of the author's $200 weekly plan limits.
- Pinned to TypeScript commit 673a5f17d713 (29 September 2026, 7.1.0-dev). All 181,711 ported Go tests pass; TanStack Query core and Hono give identical diagnostics.
- On 60 open-source projects, type checking takes about half of Go's time (geometric mean); identical output on all 60, versus 37 of 60 for bun check.
- Six-app benchmark (VS Code 3.75M lines, Sentry, Playwright, Excalidraw, TypeORM, tRPC): geometric-mean speedup over tsc 6 of 7.1x for tsc 7, 13.4x for tsc-rs, 20.9x for bun check.
- VS Code: 54.56s (tsc 6), 6.84s (tsc 7), 3.49s (tsc-rs), 1.62s (bun check).
- T3 Code with Effect diagnostics: tsc-rs 8.35s versus 21.07s for tsc 7 plus @effect/tsgo.
- Known problems: monorepo edge cases, build ordering differences in tsc -b, editor memory growing about 20 MiB per 1,000 edits.
How they got it
The author ran agent loops against the upstream Go test suite as an oracle. Benchmarks used hyperfine, median of 5 runs after 1 warmup, on an Apple M4 Pro (12 cores, 48 GB), with commit hashes listed for every app.
Limits
- Most of the README below the cost section was itself written by the LLMs; claims are reproducible but not independently audited.
- The OpenAI-versus-Anthropic comparison is one person's runs, not a controlled test; the earlier code may have helped indirectly.
- Porting with a complete test suite and a reference implementation is an easier task than building new software.
Why it matters for companies
A large, well-specified rewrite with an existing test oracle can now be done by agents at a cost below a single engineer-year. The deciding input was the test suite, not the model alone. Cost varied by more than 16x between model choices for the same goal. Companies with legacy systems should ask whether they have tests strong enough to let agents verify a rewrite, and should budget pilots per model.
Summary based on the full text. Read the original for the full argument and data.Read the original on github.com