How to route Claude Code subagents to cheaper models without losing quality

Most of the cost of an AI coding session comes from one habit: running every step on your most capable model. Reading files, grepping, and summarizing logs do not need it. Claude Code subagents let you split the work, and recent releases make the split easier to control. This guide shows a simple routing approach you can try today.

The idea: match the model to the job

Think of your session as a small team. One senior engineer (a strong model) decides what to build and reviews the result. Junior helpers (small, cheap models) fetch information and report back. The expensive model only sees the short summaries, not the raw output.

Good jobs for a small model:

• Searching the codebase and listing relevant files
• Summarizing long test or build logs
• Collecting documentation for a library
• Checking that a list of files follow a naming rule

Keep a stronger model for:

• Designing the change
• Writing or modifying core logic
• Security-sensitive review

Step 1: define a helper subagent

Subagents are markdown files in .claude/agents/. The frontmatter can name a model. A minimal research helper might look like this:

---
name: repo-scout
description: Finds files and summarizes how a feature is wired. Read-only.
model: haiku
tools: Read, Grep, Glob
---
Search the repository for what the user asks about. Return a short list of
file paths with one line each explaining why it matters. Do not edit anything.

Giving it read-only tools is a safety measure as much as a cost one: a cheap helper that cannot write cannot break anything.

Step 2: keep helper context small

Per the Claude Code changelog, version 2.1.296 added autoCompactWindow to subagent frontmatter, so a subagent can auto-compact earlier than the main conversation's window. If a helper tends to wander through many files, set a smaller window so it summarizes its own history sooner instead of dragging a huge context along.

Step 3: pin workflow agents to one model

If you run multi-agent workflows, the same release added CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL. It runs every workflow agent on one model while other subagents keep theirs. Set it in your shell before launching Claude Code when you want a whole batch of workflow agents to run on a cheaper model:

export CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL=claude-haiku-5-5
claude

The model ID above is the one the changelog lists for Claude Haiku 5.5 in version 2.1.293. Check that your account and provider expose it before relying on it.

Step 4: tune effort per task

Version 2.1.292 added an effort parameter to the Agent tool, per the Releasebot digest of recent Claude Code releases. Lower effort is a sensible default for lookup-style helpers, and higher effort for anything that must reason carefully.

Step 5: verify the split before trusting it

Cheaper is only better if the answers are as good. Run the same three tasks twice, once with the helper on your strong model and once on the small one, then compare:

1. Did it find the same files?
2. Did the summary leave out anything you needed?
3. Did total cost or time actually drop?

If a helper misses things, move it back up a tier or tighten its instructions. Keep a short note of which helpers you downgraded and why, so you can revisit the choice when models or prices change.

A quick checklist

• Read-only helpers on a small model; writers and reviewers on a strong one
• Helper outputs are short summaries, not raw dumps
• autoCompactWindow set for helpers that read a lot
• One side-by-side comparison before you commit to a downgrade