Claude Code users may have noticed it getting to work a little sooner, and Anthropic now has numbers to back that up. In a post on X at 05:34 Beijing time on October 17, the ClaudeDevs account wrote, "We keep improving performance across Claude, and you may have noticed it starts responding sooner!" The biggest gain it reported was in Claude Code's local sessions, where the wait before the first token is now about a third of what it was in August.
What Anthropic changed
The post points to two changes. "Claude Code now gzips the conversation it sends on every turn, so it reaches the model faster," ClaudeDevs wrote, adding that "cloud sessions also start up as soon as you start typing." It doesn't say how much of the improvement comes from each change, or whether other work also played a part.
The numbers, and how they were measured
The chart attached to the post compares August with October 2026, and three of its four rows use the same measure: the wait before the first token, excluding the time the model itself spends, at the 90th percentile. On that basis, Claude Code local sessions went from 4.07 seconds to 1.33 seconds, which the chart labels 3x faster. Claude Code cloud sessions went from 16.99 seconds to 6.00 seconds, or 2.8x by the chart's count, and Claude chat and Cowork went from 3.36 seconds to 2.17 seconds, or 1.5x.
The fourth row, for the Claude API, is measured differently. It shows median (p50) time to first token, and this time model time is included. That figure fell from 1.53 seconds to 1.01 seconds, which the chart also calls 1.5x faster. Because it uses a different percentile and includes the model, it can't be compared directly with the other three.
The chart's headline reads "3x faster in Claude Code, 1.5x faster in Claude," but the 3x figure belongs to local Claude Code sessions specifically. All of these are Anthropic's own measurements, and the company hasn't published more detail on how they were collected.
The changelog entry
The compression change also shows up in the Claude Code changelog. Version 2.1.300, published to npm at 22:27 Beijing time on October 14, includes this line: "Changed large requests to api.anthropic.com to use gzip compression even when telemetry is disabled; set CLAUDE_CODE_GZIP_REQUEST_BODIES=0 to turn it off." In other words, large requests to Anthropic's API are now compressed whether or not telemetry is turned on, and setting the environment variable CLAUDE_CODE_GZIP_REQUEST_BODIES to 0 switches compression off. The changelog doesn't say when someone might want to do that.
