Claude Code now gzips the conversation on every turn
Anthropic shipped a transport change, Theo says steering mid-task is where you notice it, and nobody has posted latency numbers.
The ClaudeDevs account says Claude Code now gzips the conversation it sends on every turn so it reaches the model faster, and that cloud sessions start up as soon as you start typing. The framing is performance, not capability. Nothing about the model changed in the post, only how much data has to cross the wire before the model sees your turn.
Claude Code now gzips the conversation it sends on every turn, so it reaches the model faster.
Theo says he can feel it. He describes the model starting to respond almost immediately after he presses enter, calls it almost trippy, and says he did not expect the difference to be this noticeable. That is one engineer's impression on his own machine and his own network, not a measurement. No one in either post has published latency numbers, before or after, so the size of the win is unknown.
Where it shows up
The detail worth keeping is where Theo says the change lands hardest.
"Diff is most notable with steering imo, you can see model correcting in real time."
Diff is most notable with steering imo, you can see model correcting in real time.
Steering here means interrupting an agent while it is working and redirecting it, rather than queueing a follow-up and waiting for the current run to finish. Theo has already said he stopped queueing messages in favour of correcting agents mid-run, so a round trip that returns faster hits exactly the workflow he already uses.
For working engineers, the thing underneath is that a coding agent resends conversation history on every turn, and that history grows as the session runs. Compressing it before it goes out is an upload problem, not an inference problem, which is why the effect is described as the model starting to respond sooner rather than finishing sooner. Whether you notice depends on your connection and how long your sessions get. Someone on a fast uplink with short sessions may see nothing at all.
The honest summary is a small infrastructure change, one enthusiastic report from a heavy user, and no public numbers yet. If perceived latency is what makes mid-run steering feel worth doing rather than annoying, transport work like this changes how people drive agents and not just how the terminal feels.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.