Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Claude Code subagent updates now cost 99% less

Written November 6, 2026

Background subagents used to re-read their whole conversation every 30 seconds. Haiku 5.5 now writes status from short notes. What that saves.

What changed

On November 4, Anthropic's @ClaudeDevs account posted that background subagents in Claude Code now use less of the usage limit. A running subagent updates its status line — the line that says what it is doing — every 30 seconds, and each update used to re-read the subagent's whole conversation. Now Haiku 5.5 writes the status line from short notes instead, so each update costs about 99% less.

The change targets the status line, not the subagent's work. The subagent keeps running on its own model; only the author of the status line changes. That distinction is the whole story: a small model now writes a line that used to cost a full conversation re-read, 120 times an hour per running subagent.

What it is not

The post names no Claude Code version carrying the change, and the change may be server-side — nothing in the post says to update. The public changelog lists a related fix, not this change: version 2.1.318 lists a fix for the "Now using usage credits" notice reappearing about every 30 seconds while a background subagent is running. That is the same 30-second rhythm, but a different fix, and the changelog does not name the status-line change.

It is also not a 99% cut to a bill. The 99% figure is per status update. Anthropic gave no token counts and said nothing about how much of a subagent's total usage the status line took before, so the saving per session is not known from the post. A session with no background subagents saves nothing; a session with several subagents running for hours saves the most, because that is where the 30-second updates pile up.

What it saves, in illustrative arithmetic

The arithmetic below is illustrative only, labelled as such. Anthropic gave no token counts, so no measured saving exists to report.

One update every 30 seconds is 120 updates an hour for each running subagent. At about 1% of the old cost each, those 120 updates cost roughly what 1.2 of the old updates did, if each old update cost the same. Five background subagents running for an hour would produce 600 status updates by the same arithmetic, each about 99% cheaper than before. What that is worth in tokens depends on how large each subagent's conversation had grown, because the old cost was a full re-read — and that size is exactly what Anthropic did not publish.

The practical read: the saving is per update, not per task. A long session with background subagents should draw less from the usage limit than the same session drew before, and the difference grows with the number of subagents and the length of the session. How much less, in tokens or in prompts, is not something the post lets anyone calculate.

What to check

Watch /usage totals on a long session with background subagents running, and compare against a similar session from before the change. That is the only measurement available to a reader: the per-update figure is Anthropic's, the session total is yours to observe. If the session has no background subagents, expect no difference — the status line only exists while a subagent is running.

For the wider picture of what a Claude Code session draws from a plan, see baseline usage before the shared allowance and run parallel subagents in Claude Code.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Baseline usage before the shared allowance · Claude Code advisor: Sonnet or Opus · Run parallel subagents in Claude Code