Claude Code Subagents: Status Updates Now Cost About 99% Less

What Haiku 5.5 writes every 30 seconds

Now Haiku 5.5 writes it from short notes instead, so each update costs about 99% less.

Background subagents in Claude Code keep a status line while they run. That line used to be rebuilt by rereading the subagent's whole conversation on every refresh. ClaudeDevs now has Haiku 5.5 produce the same kind of update from short notes, so the status pass no longer drags the full transcript back through the model.

The cadence did not change. A running subagent still updates its status line every 30 seconds. What changed is the input to that refresh: short notes instead of the entire conversation, which is why ClaudeDevs can say each update uses less of your usage limit.

• Old refresh

  Each 30-second status update reread the subagent's whole conversation.

• New refresh

  Haiku 5.5 writes the line from short notes rather than the full transcript.

• Who announced it

  ClaudeDevs, on November 4, 2026, with the docs living on code.claude.com.

Where the savings actually show up

A 99% cut applies to the status update itself, not to every token the subagent spends on the task. ClaudeDevs also frames the practical result at session scale: whole-session usage is about 5% lower, because the status line is only one slice of the work.

The gap matters most when a background subagent stays alive long enough to emit many refreshes. Thirty-second updates add up, and the old path paid a full-conversation read each time. Task work, tool calls, and the main conversation still sit outside this particular saving.

How this fits the subagent docs

The official Claude Code docs on code.claude.com cover creating custom subagents for task-specific workflows and tighter context management. The same docs area includes built-in subagents, a quickstart, configuration, scope, subagent files, and a frontmatter reference, alongside broader sections on agents and parallel work.

You reach the product through Claude Code, with the documentation on code.claude.com. The November 4 change does not publish a separate price list, a new plan, or a new install path. It changes the cost of an existing behavior: the background status line on a subagent that is already running.

What the announcement leaves open

ClaudeDevs states the cost result clearly and does not publish a side-by-side quality check of note-based status lines versus lines written from the full conversation. If you rely on that line to judge whether a long job is stuck, the useful open question is whether the shorter notes stay specific enough after many 30-second cycles.

There is also no new benchmark table and no stated change to how custom subagents are defined. The known limits are the ones in the announcement: about 99% less per status update, about 5% less usage across a session, and a refresh that still happens every 30 seconds.