Claude Code Rereads the Entire Conversation Every 30 Seconds Just to Update One Status Line
Anthropic changed Claude Code's background subagent status line so that Haiku 5.5 generates it from short notes, cutting the cost of a single update by about 99%. The company says sessions with background subagents use about 5% less quota, and the same quota supports roughly 6% more prompts. CLI users can get it by upgrading to v2.1.316+.
When Claude Code's background subagents are running, the status line updates every 30 seconds. Previously, each update reread the entire conversation. Now Anthropic has changed it so Haiku 5.5 writes the status line from short notes, reducing the cost of a single update by about 99%.
The two official numbers need to be read together: sessions with background subagents use about 5% less quota; with the same quota, users can get roughly 6% more prompts. The 99% reduction applies to the cost of a single status-line update, while the cost reduction for the whole session is around 5%. This distinction is important; otherwise it is easy to overestimate the effect of the optimization.
The subagents themselves still use their own models. Haiku 5.5 is only responsible for writing the status line and does not replace the subagents' work. This division of labor also explains why a single update can drop so much: the status line needs only a short description, not the full context fed back into the model.
Availability: it is already available to most Pro, Max, and Ultra users, as well as Team and Enterprise users. Users need to upgrade to v2.1.316+. Also, this change only takes effect in the CLI; the desktop app does not write this subagent status line, so there is no corresponding change there.
The comments are fairly consistent. Some ask how the 99% is achieved, some do the math and say the status line has been eating quota all along, and others joke that the entire conversation is reread every 30 seconds just to update one line of text. One commenter described it as making Haiku fill out reports. The complaint is not unfair: the status line is not core reasoning, yet it was running the way core reasoning does, and the waste was hidden in the edges.
For people building agent products, this detail is more useful than the numbers themselves. Seemingly minor features such as progress display, logs, and status synchronization can quietly pile up costs if they carry the full context every time. What Claude Code changed this time is a small interface, but it exposes a common category of overhead.
Source: ClaudeDevs' X post https://x.com/ClaudeDevs/status/2118063143167693011
发布时间: 2026-11-05 23:37