What ClaudeDevs said changed
ClaudeDevs posted on October 16, 2026, that performance work across Claude is already noticeable because responses start sooner. The note gives no millisecond figure, model name, or separate ship time. The post time, 21:34 UTC, is when the account published the note.
ClaudeDevs says Claude Code now gzips the conversation it sends on every turn, so that payload reaches the model faster. The Claude Code changelog states a narrower rule: large requests to api.anthropic.com use gzip compression even when telemetry is disabled, and CLAUDE_CODE_GZIP_REQUEST_BODIES=0 turns that compression off. The bullet gives no byte cutoff, no Claude Code version, and no measured saving. “Every turn” and “reaches the model faster” remain the account’s wording.
ClaudeDevs says cloud sessions start up as soon as you start typing. The changelog notes compared here run through the line on the interactive session’s WebSearch budget, near the end of that long page. In that stretch the cloud-session items are fixes — startup error text, rewind, diffs, routine runs — and there is no sentence that a session begins when typing starts. The page continues after that line, so a sentence only in the remaining stretch sits outside this comparison. Mechanism, plan, and any time saved stay unspecified.
On the API, ClaudeDevs says the Messages API now does less work before a request reaches the model, and that this lowered time to first token. The same note says requests are no longer buffered on the way in, on top of a long list of smaller infrastructure improvements the post leaves unnamed. Those API claims stay with ClaudeDevs, with no model, request class, or millisecond figure attached.
Who the note is for
Claude Code users who send requests to api.anthropic.com are the group with a documented client change. That gzip rule covers large requests to that host. The post’s “every turn” line would cover smaller conversations as well. The changelog leaves “large” undefined.
Cloud-session users are the audience ClaudeDevs names for the typing startup. The post names no plan, region, or client. Messages API callers are the audience for the time-to-first-token claim and for unbuffered ingress. The post describes both as already in place. It mentions no opt-in, price, or credit terms.
How to keep or turn off gzip
ClaudeDevs presents the changes as already shipping. The post gives no install step.
For Claude Code, the documented control is the environment variable in that gzip rule. Leave CLAUDE_CODE_GZIP_REQUEST_BODIES unset to keep gzip on those large requests. Set CLAUDE_CODE_GZIP_REQUEST_BODIES=0 to turn the compression off. The changelog says to run claude --version to see the installed version. The gzip bullet itself names no version, so a version string alone does not show whether this change is in your build.
Cloud-session startup and the Messages API path have no command, flag, or opt-out in the post. A faster-feeling reply matches the account’s qualitative claim. The post does not say how to measure it.
What is still open
ClaudeDevs gives no latency number, percentile, model, plan, or region. “Starts responding sooner” and “lowered time to first token” are the account’s descriptions.
October 16, 2026, at 21:34 UTC is the publication time of the note. It is a single timestamp for the whole post, so it is not a confirmed ship time for gzip, cloud sessions, and the Messages API separately.
The gzip bullet leaves the size threshold, other API hosts, a version, and any faster time to the model unstated.
Cloud sessions starting when you type, the Messages API pre-work cut, and unbuffered ingress are ClaudeDevs’ account. In the changelog stretch compared above, those three are unstated.
Source
Also from
Sources and editorial approach
This brief relies on the ClaudeDevs post published October 16, 2026, and on the Claude Code changelog at code.claude.com as checked that same evening. The gzip sentence is the changelog’s. The comparison covered the release notes through the WebSearch budget line near the end of that long page; the last stretch of the page was not included, so a line only in that stretch would have been missed. No Messages API reference was opened. No Claude Code session, cloud session, or API call was timed.