Anthropic cuts Claude's wait before the first token: Claude Code local sessions 3x faster (4.07 s to 1.33 s, p90), cloud sessions 2.8x, chat/Cowork and API 1.5x, helped by gzipping each turn's conversation

Why it matters

【Key points】

Anthropic's developer account @ClaudeDevs posted on Oct 16 at 21:34 UTC that Claude products now start responding noticeably sooner, with a chart comparing August and October 2026. The post has passed 2 million views on X.

The first three rows measure the wait before the first token, excluding model time, at p90. Claude Code local sessions went from 4.07 s to 1.33 s (about 3x faster). Claude Code cloud sessions went from 16.99 s to 6.00 s (2.8x). Claude chat and Cowork went from 3.36 s to 2.17 s (1.5x). The Claude API row uses median (p50) time to first token including model time: 1.53 s to 1.01 s (1.5x).

Anthropic named two changes. Claude Code resends the whole conversation every turn and now gzips it first, so it reaches the model faster. Cloud sessions also start booting as soon as the user starts typing. The post does not detail the other improvements.

【Verifiable】

@ClaudeDevs post with chart: https://x.com/ClaudeDevs/status/2111209208012210409

Claude Code changelog: https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md

【Our read】

1. The speed-up is engineering work; the models did not change. Claude Code calls are stateless and carry the full context every turn, so on long sessions the upload itself is a large share of the wait, and compressing it saves whole seconds. The viral line "gzip from 1992 made Claude Code 3x faster" overstates it: the 3x figure combines several changes and counts only the wait outside model time. It does not mean responses are 3x faster end to end.

2. The figures are Anthropic's own p90/p50 metrics. No methodology is published and nobody has independently reproduced them, so treat them as vendor numbers.

3. Heavy users will feel the cloud-session drop from about 17 s to 6 s the most. In the same week, OpenAI Codex users are complaining about quotas draining fast. Competition between coding tools is moving beyond benchmark scores to latency and usage limits.