Claude Code 默认启用 gzip：首字延迟最多快 3 倍

Aakash Gupta @aakashgupta原推发布于 2026.10.18

在 X 上打开原推

Claude Code just got up to 3x faster at starting its replies, and part of the fix is a file-shrinking format from 1992 that your browser already uses on almost every web page you open. Here's the weird part. Every single time you hit enter, your laptop re-uploads the entire conversation to the AI from scratch. In a long session that can be a whole megabyte per message, and until recently none of it was compressed. AI models don't remember anything between messages. The conversation lives on your machine, so every turn Claude Code bundles up everything, your prompts plus every file it opened and every test it ran, and ships the whole pile again. Anthropic's servers cache the processed version so the model doesn't have to re-read it, but the raw bytes still have to make the trip up your connection every time. And upload is the slow side of most internet plans. On a 5 Mbps uplink, which is about what a lot of hotel Wi-Fi gives you, a 1 MB request takes 1.6 seconds just to leave your laptop. An agent fixing one bug makes a separate round trip every time it opens a file or runs the tests, and each one carries the full conversation with it. A developer actually caught this back on August 28. They pointed Claude Code at a logging proxy and watched the requests go out completely uncompressed, one of them 594 KB on its own. The API already accepted gzip, and the app just never sent it. Their workaround was a 170-line Python script that shrank real traffic about 2.1x and saved 185 MB of uploads from one machine over three hours of normal work. Claude Code gzips by default now. For the slowest 10% of requests, the wait before the first word shows up (not counting the model's own thinking) dropped from 4.07 seconds to 1.33. Cloud sessions went from almost 17 seconds down to 6. Anthropic hasn't said how much of that came from compression versus the other changes they shipped alongside it. The slowest 10% is exactly where the bad connections live, though, and cutting a 1.6-second upload in half hands you back most of a second on every round trip, all day long. Everybody's spending billions to make the models think faster, and it turns out a lot of developers were sitting there waiting on a megabyte of text squeezing out through the upload side of their hotel Wi-Fi.

这条是引用着另一条写的：x.com/2024518793679294464/status/211120920801221…