Plan with a frontier model, code with a fast one
Teams that want frontier-quality plans without paying frontier prices for every edit.
| Cost | Per token: the plan is a small share of the tokens, the edits run on a model about 40 times cheaper per token |
|---|
What is in it
| Role | Part | Why |
|---|---|---|
| Coding agent | Claude Code Included in paid Claude plans: Pro $20/month ($17/month billed annually), Max from... |
Plan mode for the approach, subagents for the work. |
| Planner | Claude Opus 5.5 $4 / $20 per 1M tokens |
Writes the plan and reviews the result. |
| Worker | Claude Haiku 5.5 $0.1 / $0.5 per 1M tokens |
Set model: haiku on subagents for routine edits and test runs. |
| Spend tracking | LiteLLM The SDK and proxy are open source and free. An enterprise edition with SSO, audit logs... |
Optional gateway that shows cost per key and per model. |
Set it up
npm install -g @anthropic-ai/claude-code- Start complex work in plan mode (Shift+Tab), approve the plan, then let subagents do the steps.
- In each subagent's configuration set model: haiku for simple, repetitive jobs.
Trade-offs
A weak worker model can miss intent on subtle changes; keep tests as the check and review the diff.
Comments
- Using this stack? Share what worked and what did not.
Other stacks
Claude Code on a $20 plan, without hitting the limit Open-source agent with a cheap model, paid per token A coding agent with guardrails for a team