Briefs · Taipei time (GMT+8)

Claude responds faster: time to first token is 3x quicker locally in Claude Code and 1.5x quicker on the API

一句重點：Most of the speedup comes from transport and preprocessing. Every turn of a Claude Code conversation is now gzip-compressed before it goes out (@ClaudeDevs), and the API does less work on a request before handing it to the model (@ClaudeDevs thread).

發生什麼事

Claude Code's cloud sessions have had their startup timing changed too: a session gets launched ahead of time, at the moment the user begins typing (@ClaudeDevs). On the API side, incoming requests are no longer buffered first, and a long list of smaller infrastructure improvements went in as well (@ClaudeDevs thread).

The attached chart compares p90 wait times, excluding model time, from August to October 2026. Local Claude Code dropped from 4.07 to 1.33 seconds, cloud from 16.99 to 6.00 seconds, and chat and Cowork from 3.36 to 2.17 seconds. For the API, p50 including model time went from 1.53 to 1.01 seconds (@ClaudeDevs).

01Sources

  • @ClaudeDevs main post: Anthropic's official developer account, 10/16 21:34:34 UTC (05:34 on 10/17, Taipei time), about 1,400 likes / 88 replies; Claude Code sends conversations gzip-compressed, cloud sessions start once typing begins, chart comparing wait times by product
  • @ClaudeDevs thread, second post: posted at the same time; less preprocessing in the Messages API, requests no longer buffered, other infrastructure improvements

More briefs →