Claude Code gzips each turn and pre-starts cloud sessions

Anthropic's ClaudeDevs account says Claude Code now starts responding sooner and points to two changes outside the model itself. First, Claude Code now gzips the conversation it sends on every turn. Each turn ships the conversation to the model, so in long sessions the upload can be large, and compressing it means fewer bytes travel before the model can begin. Second, cloud sessions, where Claude Code runs on remote infrastructure rather than the user's machine, now begin starting up as soon as the user starts typing, so the setup work overlaps with writing the prompt instead of following it.

The chart attached to the post measures the wait before the first token appears, excluding the time the model spends working, at the 90th percentile, meaning the slower end of typical requests, from August to October 2026. On that measure, local Claude Code sessions went from 4.07 to 1.33 seconds, about three times faster, and cloud sessions from 16.99 to 6.00 seconds, 2.8 times faster. The Claude chat app and Cowork improved from 3.36 to 2.17 seconds, and the Claude API's median time to first token, which does include model time, fell from 1.53 to 1.01 seconds. The post presents these as the result of ongoing performance work and does not say how much of each gain came from compression, early session start or other changes.

This continues earlier startup work: in August, Anthropic said the Claude Code command-line tool no longer waits for the sandbox and MCP servers before letting users type.