Author reports a never-read Rust port of the TypeScript compiler, at over $420,000 in tokens on OpenAI models and about $24,000 on Claude Opus 5.5

The README says the project ports Microsoft's native TypeScript compiler, checker and language server from Go to Rust. The author writes: "I've never read a line of this code." The README reports more than $400,000 of API-priced tokens on GPT-5.6 Sol and GPT 6 Astra, spent over multiple months of /goal loops, producing more than 1.3 million lines of Rust that "never got past like 84% compat". The author then tried Opus 5.5, which "started from scratch. It got further than Astra in 1/10th the time." The stated total for that run is about $24,047 of API spend over two weeks, which the README says came to between 925% and 983% of a $200 plan's weekly limits. It also says the whole job could "probably" have been done for about $20,000, which is the author's estimate.

The README makes testable claims and ships the benchmark scripts. It reports that all 181,711 ported Go tests pass, that type checking takes about half of Go's time (geometric mean) across 60 open-source projects, and that command-line output matches Go's on 120 repositories apart from the listed problems. On six apps the table shows a geometric-mean 1.61x speedup over tsc 7. These are the author's measurements on one Apple M4 Pro machine, and we did not run them. The listed problems include tsc -b --watch stopping with exit code 70 after some edits, and editor memory that grows during long sessions. A section the README calls "The Slop Line" marks everything below it as written by the author's models.

Why it matters: The useful part is the verification setup, not the headline. With 1.3 million lines nobody reads, what the loops had to lean on was an oracle: Go's own test baselines, run against a pinned upstream revision. If you plan a long unattended loop, check whether you have an equivalent oracle before you start. The cost split is one author's account of one job, so read it as a data point on restarting with a different model, not as a general price comparison.