Claude Code cuts first-token latency threefold with gzip

Anthropic says Claude Code’s first-token latency fell sharply between August and October, mostly from infrastructure changes rather than a faster model. For local sessions, the p90 time before the first reply token dropped from 4.07 seconds to 1.33 seconds. Cloud sessions fell from 16.99 seconds to 6.00 seconds, and claude.ai’s web and Cowork experiences improved from 3.36 seconds to 2.17 seconds.

Most of the local gain comes from gzip-compressing per-turn conversation payloads before they reach the API, a change that shipped in Claude Code 2.1.300 on 14 October. It helps most with long contexts: pasted logs, large files, sprawling back-and-forth, where the uncompressed payload was doing the most damage. Cloud sessions improved separately, by starting the remote session as soon as you begin typing rather than after you hit send.

Why it matters: latency is one of the main things that makes an agentic coding tool feel usable rather than sluggish, especially if you regularly pipe large files or long sessions through Claude Code. The fix is already live and opt-out, not opt-in.

export CLAUDE_CODE_GZIP_REQUEST_BODIES=0  # disable if you need to

The caveat: the headline numbers come from Anthropic’s own measurements, posted through its ClaudeDevs account rather than an independent benchmark, and cover p90 latency specifically, not every request.