Anthropic / Wikimedia Commons (Public domain)
Gzip compression and earlier cloud session startup drove the latency drop across local, cloud, web and API surfaces
Anthropic has made Claude Code noticeably less sluggish. First-token latency for local sessions fell from 4.07 seconds to 1.33 seconds between August and October 2026, an improvement of nearly 3x.
The fix was not a new model or an exotic breakthrough. It came from compressing conversation data and starting cloud sessions before users finish typing.
Where the seconds went #
The update was announced on October 17, 2026, through the ClaudeDevs account. It covers several Claude products, with the headline gains in Claude Code.
The key metric is first-token p90 latency. First-token latency is how long you wait before the model starts producing any reply at all. The p90 label means 90% of requests came back faster than that figure.
For local Claude Code sessions, that p90 number dropped from 4.07 seconds to 1.33 seconds. Cloud sessions saw an even larger absolute cut, going from 16.99 seconds to 6 seconds, roughly 2.8x faster. The gains spread beyond the coding tool. Claude’s web and Cowork experiences improved by about 1.5x, with response times falling from 3.36 seconds to 2.17 seconds.
The API also got quicker. First-token response time there went from 1.53 seconds to 1.01 seconds.
The two tricks behind the speedup #
Anthropic started gzip-compressing those per-turn conversation payloads before sending them to the model. Gzip is a long-established compression format that shrinks data so it moves across the network faster. The benefit is biggest for long contexts, where developers often paste in large files, logs, and sprawling back-and-forth exchanges.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The second change targets cloud sessions. Previously, a cloud session would spin up only after a user submitted a prompt. Now the session begins warming up the moment someone starts typing.
That pre-start mechanism helps explain why cloud latency saw the most dramatic drop. The startup cost does not disappear, it simply gets hidden inside the time a person spends writing their request.
Some reactions to the announcement pointed out that gzip is hardly new technology, though most user feedback focused on the speed gains themselves.
Part of a broader performance push #
The announcement follows an internal sprint the previous month in which more than 3,000 updates produced similar performance improvements across various Claude products. The Claude Code numbers look like one visible result of that wider effort.
The timing also landed right after news that Claude generates approximately 90% of its own code.
What this means for developers and the AI coding market #
For developers, latency is not a vanity metric. Coding assistants are used in rapid, iterative loops: ask, read, tweak, ask again. A few seconds saved per turn compounds across hundreds of turns in a working day. The cloud improvement may matter most here. Going from 16.99 seconds to 6 seconds changes whether a developer stays in flow or drifts off to another tab while waiting.
By showing gains at the p90 level across local, cloud, web, Cowork and the API, Anthropic is signaling that it sees speed as a product feature across its whole lineup rather than a single tool.
There are limits worth keeping in mind. A 6-second wait for cloud sessions is still noticeable, and the pre-start approach means compute spins up for prompts some users may never actually send.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our