One Million Lines in 11 Days: AI Rewrote the Entire Bun Runtime in Rust
The JavaScript tool Bun has been fully rewritten from Zig to Rust — and Claude Fable 5 did most of the work, running 64 instances in parallel for 11 days at a cost of $165,000.
Bun, the popular JavaScript runtime known for its speed, has been completely rewritten in a new programming language — and almost none of the code was typed by a human. Creator Jarred Sumner says about 64 instances of Anthropic’s Claude Fable 5 ran in parallel for 11 days, producing over a million lines of Rust code to replace the original Zig codebase. The API bill: roughly $165,000. A human team, Sumner estimates, would have needed about a year.
The reason for the switch was reliability, not speed. Zig, the language Bun was originally built in, kept producing memory errors and crashes that were hard to stamp out for good. Rust — a language famous for catching exactly this class of bug at compile time, before the program ever runs — promised fewer of those late-night surprises. The rewritten version, Bun v1.4.0, is out as a canary release (an early version for people willing to test), fixes 128 known bugs, and runs about 2 to 5 percent faster.
What’s behind this? Rewriting a large codebase in another language is one of the most dreaded jobs in software — enormously expensive, boring, and risky, which is why most companies simply never do it. That’s exactly why AI labs love these stories: they’re proof that AI agents can now handle sustained, verifiable work at a scale no human team would attempt on that timeline. One honest caveat belongs in the frame, though: Anthropic acquired Bun and Sumner’s team in December 2025. This rewrite was done by Anthropic’s model, on Anthropic’s product, paid with what is effectively Anthropic’s own money — a genuine technical feat, but also a showcase staged under ideal conditions. Reportedly the million lines still needed human review and the canary label means it’s not yet considered fully stable.
What this means for you: If you don’t write code, take this as a data point on speed: work measured in team-years is starting to compress into weeks, and that pattern won’t stay confined to programming. If you do write code, the interesting part is the method — many parallel agents, an expensive model, and a compiler strict enough to catch machine mistakes. That last bit matters: Rust’s strictness acted as a safety net for AI-generated code, which hints at how these large automated rewrites will be made trustworthy. And if you use Bun: maybe let the canary fly for a few weeks first.
Sources
Source: https://bun.com/blog/bun-in-rust
Light to Ultra: Making Sense of GPT-5.6's Seven Thinking Speeds
GPT-5.6 Sol asks you to choose how hard it should think. An OpenAI staffer has now explained what each reasoning level is for — here's the plain-language version.
