AI Coding Weekly

Bun AoT starts Claude Code 24% faster, says Jarred Sumner

The experimental build also cuts the Claude Code install from 430 MB to 304 MB, with a nearly 2x larger binary.

Jarred Sumner posted numbers from an early preview of experimental Bun plus JSC with ahead-of-time compilation, and the test subject is the Claude Code CLI. His figures are 24% faster startup, 48% lower memory, 46% lower CPU, and nearly 2x larger binary size.

Ahead-of-time compilation means the JavaScript is compiled to machine code before the program runs, instead of being compiled by the engine while it runs. The trade shows up directly in Sumner's own numbers. You pay in binary size and you get back startup time, memory and CPU.

The size numbers cut both ways

In a follow-up post Sumner says Claude Code is now 304 MB instead of 430 MB. That sits oddly next to the nearly 2x larger binary from the first post, and he does not reconcile the two in either message, so read them as what they are, separate measurements from an experiment in progress.

Jarred Sumner
@jarredsumner
X
Other workloads are different though, TypeScript 6 runs faster in Bun AoT than in Bun JIT now
Oct 4, 2026 · View on X

He is also explicit that the work is unfinished. The typed part, he writes, is still pretty half baked, and he thinks it is leaving a 2x to 10x on the table. He adds that other workloads behave differently, and that TypeScript 6 now runs faster in Bun AoT than in Bun JIT.

Why startup is the number to watch

For working engineers the interesting metric here is not the headline percentage, it is which metric got improved. Agent CLIs are not long running servers. They get launched over and over through the day, often several at once, and every launch pays the full cost of booting a JavaScript runtime before any work happens. Startup time and resident memory are the part of that cost a developer actually feels, especially anyone running multiple agent sessions on one laptop.

The caveats are the usual ones for a preview. These are one author's numbers on one CLI, from a build he describes as experimental, with no third party reproduction posted and no release date given. Nobody has published independent measurements yet.

Get the next one by email

Coding with AI, read daily so you do not have to. The experiments, the receipts and the arguments from engineers shipping real software. Not a changelog.