Skip to main content
Official reference — The official costs docs and model configuration docs cover pricing, caching, and the CLAUDE_CODE_SUBAGENT_MODEL setting. This page adds the opinionated routing strategy.

​
The Cost Reality

The overhead comes from each agent loading project context independently, every message consuming tokens in both sender’s and receiver’s context, and broadcasts multiplying cost by teammate count.

​
Model Routing Strategy

The most impactful cost optimization: use the cheapest model that can do the job at each stage. Example pipeline:

​
The opusplan Pattern

Official reference — For model aliases, /model switching, and CLAUDE_CODE_SUBAGENT_MODEL, see the official model configuration docs. This section covers the routing strategy.
The highest-leverage cost optimization: Opus plans, Sonnet executes. Use Opus for architecture decisions and coordination, then route implementation to Sonnet agents. This cuts 60-80% of token costs while maintaining quality where it matters. The key insight: model choice is a per-task decision, not a per-session one. Switch models as the work shifts from reasoning to execution.

​
Practical Tips

​
Minimize context pollution

  • Verbose output degrades agent performance and burns tokens. Keep agent prompts focused.
  • Use the Explore agent (Haiku, read-only) for broad searches instead of having your main Opus session read dozens of files.

​
Speed up test suites

  • Include --fast options for test suites in your CLAUDE.md — agents waste time running full suites when a subset would suffice.
  • Route test-running agents to Sonnet; they don’t need Opus-level reasoning.

​
Structure CLAUDE.md for multi-agent

  • Include module boundaries so agents know which files they own.
  • List verification commands so every agent can self-check.
  • Mark shared files (e.g., package.json, tsconfig.json) as “coordinate before editing.”

​
Use /batch for independent work

/batch is dramatically cheaper than Agent Teams for work where agents don’t need to communicate.

​
Reference: Real-World Scale

For perspective: Anthropic’s C compiler project — 16 agents over 2,000 sessions across two weeks — consumed approximately 2 billion input tokens and 140 million output tokens. Key takeaways:
  • Write extremely high-quality tests — agents solve whatever the verifier validates
  • Minimize verbose output — context pollution degrades performance
  • Maintain README/progress files so each fresh agent gets orientation context quickly
  • The constraint is always review bandwidth, not compute