Claude Code Cost Control: Agent Teams vs Subagents, Model Tiering, and the /clear Trick

TL;DR: Three levers cut Claude Code agent spend 40-50%: tier models by task role, run /clear between unrelated tasks to reset stale context, and choose subagents over agent teams when agents don't need to communicate mid-run.

Running multi-agent Claude Code workflows can get expensive fast. Three techniques that consistently move the needle:

1. Tier models by task, not by preference

Opus for decisions; Sonnet for implementation; Haiku for mechanical tasks (formatting, linting, summarizing). Lock models in your subagent YAML so the choice is enforced rather than left to each developer:

# .claude/agents/code-reviewer.md
---
model: claude-sonnet-5-20260629   # Sonnet, not Opus
tools:
  - Read
  - Grep
---
Review changed files for bugs. Return findings as JSON.

One team running tiered models costs ~40% less than all-Opus with minimal capability loss on worker tasks.

2. Run /clear between unrelated tasks

Agents accumulate conversation history across a session. Every turn re-sends that history — a 30-turn session sends 30× the tokens of turn 1. /clear resets the context window mid-session without ending it. Use it before switching to a new file or a new feature, and cut per-message token cost by 30–50%.

3. Choose subagents over agent teams when agents don't need to talk

Agent teams (experimental; enable with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=true) are powerful but expensive: each teammate maintains its own full context window, making them ~7× more token-intensive than a standard session. They're worth the cost when teammates need to coordinate mid-task — one agent flagging an API change that another agent's code must handle.

For independent parallel work — each agent reviewing a separate file, scanning a different module — use subagents instead. The orchestrator manages context; workers stay lightweight.

Rule of thumb: Independent parallel tasks → subagents. Mid-run coordination between agents → agent teams.

Track per-session usage with /usage and set org-level spend limits from the costs page.

Sources: Manage costs effectively — Claude Code Docs · Orchestrate teams — Claude Code Docs · Create custom subagents — Claude Code Docs