Claude Code Subagents Now Use Less of Your Limit

Claude Code subagents now take less of your usage limit. Anthropic moved the status line on background subagents from a full conversation re-read to Haiku 5.5 working from short notes. Each update costs about 99% less, and sessions with background subagents used about 5% less of the limit. Update to v2.1.316 or later to get it.

Try Apidog today

That's the whole story in four sentences. The rest of this post covers who gets the change, what the two percentages mean, what stays unknown, and which settings cut subagent usage further.

What changed in Claude Code subagents

A background subagent is a helper that runs alongside your main session while you keep working. In the terminal, it shows a short status line so you can see what it's doing, and that line refreshes every 30 seconds.

Anthropic announced the change on X on November 5, 2026:

• Before: each 30-second update re-read the subagent's whole conversation.
• Now: Haiku 5.5 writes the update from short notes.
• Unchanged: the subagent keeps its own model. If it runs on Opus 5.5 or Sonnet 5.5, it still does. Only the status line moved to Haiku 5.5.

You don't need to change a setting or trade away quality on the work itself. The subagent's output comes from the same model as before.

Why a subagent status line was expensive

A status line was expensive because a one-line summary was paying for a full read of a long conversation. The output was tiny, the input wasn't, and the job repeated every 30 seconds for as long as the subagent ran.

The arithmetic shows how fast that adds up. At one update every 30 seconds, a subagent running for an hour produces 120 updates. Three running for an hour produce 360.

Conversation length matters too. The Claude Code cost docs say token costs scale with context size, and a subagent's conversation grows as it reads files and runs tools. So the old design likely charged the most when a long-running subagent had the most context. That's our reading of the docs, since Anthropic hasn't published a cost curve for it.

Don't confuse this with the custom status line you set up with /statusline. The status line docs say that one runs locally and doesn't consume API tokens. The change here covers the status line Claude Code generates for each running subagent.

The numbers: 99% per update, about 5% per session

The two figures measure different things, and only one of them describes your limit.

Figure	What it measures	What it isn't
About 99% less	The cost of one status update	A saving on your bill or your limit
About 5% less	Limit used by sessions with background subagents	A saving on sessions without them
About 6% more	Prompts you get from the same limit, per Anthropic	A rise in the limit itself

If you remember one number, make it 5%. A 99% cut to something that was a small part of the session gives you a small saving on the session. Subagents still do their real work on their own model, and that work is where most of the usage goes.

Anthropic didn't say how much of the limit status updates used before. The two figures together suggest roughly a twentieth of those sessions, but that's an inference from rounded numbers, not a published figure.

Your own saving will also vary. About 5% is what Anthropic reports across sessions with background subagents. It didn't publish a range, and a session with one short subagent has less to gain than a session with several long ones.

Who gets the lower subagent usage

Anthropic says the change is live for most individual plans and all organization plans, in the CLI only.

Where you use Claude Code	Status
Pro, Max, Ultra	Live for most users, not all
Team, Enterprise	Live for everyone
CLI	Covered, on v2.1.316 or later
Desktop app	Nothing to change: it doesn't write this status line
API key, Bedrock, Vertex, Foundry	Not stated

The desktop detail comes from a reply in the same thread by Lydia Hallie of Anthropic, who said the desktop app doesn't write this subagent status line, so there was nothing to change there.

Two gaps are worth knowing about:

• "Most" has no public test. The thread and docs don't name a flag or indicator that shows whether your Pro, Max, or Ultra account has it. Being on the right version is necessary, but it doesn't prove the change is live for you.
• Pay-per-token users have no answer yet. Anthropic framed this around plan limits and didn't address API-key or cloud provider setups. The changelog for 2.1.318 does change subagent progress summaries on Bedrock, Vertex, Foundry, or with telemetry off: on a cold prompt cache, one can wait up to 5 minutes for the subagent's pending response. That shows those setups get their own handling, not that they get this saving.

The 2.1.316 changelog entries we read don't mention the status line either, so the thread is the only source for the numbers and the version.

How to update Claude Code and check your version

Three commands cover it:

claude --version   # needs to show 2.1.316 or later
claude update      # installs a newer version if one exists
claude --version   # confirm

claude update also works for package manager installs, where it updates through that package manager. The newest version listed in the changelog on November 5 was 2.1.318.

To see what subagents use, run /usage inside a session. On Pro, Max, Ultra, Team, and Enterprise plans, it attributes recent usage to skills, subagents, plugins, and MCP servers as percentages. Press d or w to switch between 24 hours and 7 days. The figures are approximate and come from local history on that machine, so treat a before and after comparison as a rough guide. If you're tracking limits closely, our Claude Pro and Max limits guide explains how they work.

Other ways to spend less on subagents

The status line fix is automatic, but it's about 5%. The docs are plain about the rest: each subagent sends its own requests, and they count toward the same limit as your main conversation. These settings have more room to help.

Pick a smaller model. A subagent inherits your session's model unless told otherwise, so a switch to Opus also applies to those subagents. Set model: haiku in a subagent definition for simple jobs. Our guide to Claude Haiku 5.5 in Claude Code covers that setup.

To set a default for all of them:

{
  "env": {
    "CLAUDE_CODE_SUBAGENT_MODEL": "haiku",
    "CLAUDE_CODE_SUBAGENT_MODEL_FORCE": "1"
  }
}

The first variable alone is only a default. A subagent definition or a model Claude passes still wins, and it doesn't change the built-in Explore and Plan subagents. The second forces one model onto every subagent.

Lower the effort. Version 2.1.316 added CLAUDE_CODE_SUBAGENT_EFFORT, which sets the effort level for subagents when nothing else names one, instead of inheriting the session's. It's in the changelog but wasn't in the environment variable reference on November 5, so check accepted values before you rely on it. For one subagent, set the documented effort field in its definition (low, medium, high, xhigh, or max, depending on the model).

Run fewer at once. The default cap is 20 running subagents per session. Set CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS to a lower whole number if you want a hard ceiling. Run /tasks to see what's running, and which model and effort each subagent uses.

Keep the brief narrow. The cost docs note that vague requests trigger broad scanning. A subagent told to check one endpoint reads less than one told to review the API. Our guide to creating custom Claude Code subagents walks through the full definition format.

Run API tests in a background subagent with Apidog CLI

API test runs suit a background subagent: they take time, produce verbose output, and you only need the failures back. Apidog CLI lets the subagent run a saved test scenario while your main session keeps coding, and the raw output stays out of your main context.

Set it up once:

npm install -g apidog-cli@latest
apidog login --with-token <YOUR_TOKEN>
npx -y skills add https://apidog.com   # pick Claude Code when prompted

Then add a small subagent at .claude/agents/api-tester.md:

---
name: api-tester
description: Runs Apidog test scenarios and reports only the failures.
tools: Bash, Read
model: haiku
background: true
---
Run the Apidog test scenario you are given. Report which steps failed
and why. Don't paste the full output.

Now give Claude Code one [REDACTED PROMPT]
Use the api-tester subagent to run Apidog test scenario in
project against environment . Keep working on
the checkout refactor. When it finishes, tell me only what failed.

The subagent runs commands like these:

bash
apidog test-scenario list --project
apidog test-scenario run --project -e -r cli,json --out-dir apidog-reports

Permission prompts from a background subagent show up in your main session, naming the subagent that's asking. For long runs, 2.1.318 also moves a shell command to the background after 3 minutes, so the prompt cache doesn't expire while it runs.

The same `run` command works in CI with `-r junit`, as the [Apidog CLI CI/CD guide](http://apidog.com/blog/apidog-cli-ci-cd-pipeline?utm_source=dev.to&utm_medium=wanda&utm_content=n8n-post-automation) shows. For the general setup, see [Apidog CLI in Claude Code](http://apidog.com/blog/apidog-cli-in-claude-code?utm_source=dev.to&utm_medium=wanda&utm_content=n8n-post-automation). Apidog's Free plan covers up to 4 users.

## FAQ

**Do Claude Code subagents now cost 99% less?** No. The 99% applies to each status update. Anthropic says sessions with background subagents used about 5% less of the limit.

**Does my subagent now run on Haiku 5.5?** No. The subagent keeps its own model. Only the status line moved to Haiku 5.5.

**What version do I need?** v2.1.316 or later. Run `claude --version` to check and `claude update` to upgrade.

**Does this apply to the Claude desktop app?** No, because there was nothing to change. Anthropic's Lydia Hallie said the desktop app doesn't write this subagent status line.

**Is it live on my Pro plan?** Probably, but not certainly. Anthropic says most Pro, Max, and Ultra users have it, and everyone on Team and Enterprise. The thread and docs don't name a way to check your account.

**Does this help if I use an API key or Bedrock?** Anthropic hasn't said. The announcement talks about usage limits and doesn't cover API-key, Bedrock, or Vertex usage.