How to reduce Claude Code costs
Practical ways to cut Claude Code token use and spend: match the model to the job, lower effort, clear context, keep the cache warm, and route each prompt.
In short
Claude Code costs come from three things: the model, the effort and the context each request carries. Sonnet handles most coding for less than Opus, and Haiku answers quick questions for a fraction of either. Lower effort cuts thinking tokens, which bill as output tokens. Clearing the chat between unrelated tasks stops Claude Code from re-sending old context with every request. A model switch in a warm chat rewrites the whole conversation into the new model's prompt cache, so switch at a natural break, such as after /clear. Keep CLAUDE.md short, because it loads in every session, and turn off MCP servers you do not use. Doing all of this by hand on every prompt is tedious, so a router can do it for you. In our benchmark of five mixed sessions, AutoMode cost $2.01 against $4.14 for Claude Code on Opus, 51% less, with all 20 steps passing their hidden tests.
Claude Code costs depend mostly on three things: the model, the effort, and how much context each request carries. Anthropic's own docs say token costs scale with context size, Sonnet handles most coding tasks and costs less than Opus, and lowering effort cuts thinking tokens on simple tasks. In our benchmark, routing each prompt to the right model and effort cost 51% less than running Claude Code on Opus, at the same pass rate.
1. Check where your usage goes
Run /usage. On a subscription plan it shows what counts against your limits, with recommendations for anything behind 10% or more of recent usage. It also reports how much of your input came from the prompt cache.
2. Match the model to the job
Compare Haiku, Sonnet and Opus. Anthropic's guidance is to use Sonnet for most coding and keep Opus for complex architecture and multi-step reasoning. Use /model to switch, and set model: haiku for simple subagent tasks. Quick factual questions rarely need more than Haiku: in our runs, Haiku answered ten questions for $0.08, against $0.50 on Opus.
3. Lower the effort on simple tasks
Thinking tokens bill as output tokens. See what each effort level does. For simple work, lower the effort with /effort or in /model. Raise it again for a hard problem.
4. Clear between unrelated tasks
Claude Code sends your whole conversation with every request. Run /clear when you switch to unrelated work. It costs nothing, and /resume brings an old session back.
5. Switch models at a natural break
Each model keeps its own prompt cache. A switch in the middle of a long chat rewrites the whole chat to the new model's cache. Switch at the start of a session or after /clear. Read when a model switch pays off.
6. Keep the cache warm
Your first message after a break longer than the cache lifetime reprocesses your full context. If you step away from a large session, expect that cost when you return.
7. Trim what loads every time
CLAUDE.md loads at the start of every session. Anthropic suggests keeping it under 200 lines and moving specialized instructions into skills, which load only when needed. Disable MCP servers you do not use with /mcp.
8. Write specific prompts
"Add input validation to the login function in auth.ts" costs less than "improve this codebase", because Claude reads fewer files. Plan mode helps on complex work, because it catches a wrong direction before the edits.
9. Route each prompt automatically
Doing steps 2, 3 and 5 by hand on every prompt is tedious, so most people leave Claude Code on one model. A router does it for you. AutoMode picks the model and effort for each prompt, checks the cache cost before a switch, and sets a quality floor for risky work.
| 5 mixed sessions, 20 steps | Passed | Questions | Coding tasks | Total | vs Opus |
|---|---|---|---|---|---|
| AutoMode | 20/20 | $0.08 | $1.93 | $2.01 | −51% |
| Plain Sonnet high | 20/20 | $0.24 | $1.90 | $2.14 | −48% |
| effortless, on Opus | 20/20 | $0.43 | $3.44 | $3.87 | −7% |
| Plain Opus (Claude Code default) | 20/20 | $0.50 | $3.64 | $4.14 | — |
Five mixed sessions, list-price equivalents, one repetition per arm. Read the full benchmark and its limits.
Sources
- Anthropic: Manage costs effectively, read on 2026-11-05.
- AutoMode: v6 benchmark report and rows.
Install AutoMode
Paste these into Claude Code, one at a time, then run /reload-plugins. It is free, needs no key and no server, and starts on its own.
/plugin marketplace add mralexivy/AutoMode
/plugin install automode@automodeLast updated on
Cost calculator
Estimate your monthly Claude Code spend on Opus, Sonnet high or effortless, and what AutoMode would cost, from per-step costs measured on real sessions.
Claude Code effort levels explained
What each Claude Code effort level does, low to max, the default for each model, how to set it with /effort, settings or env, and which level fits a task.