Optimize cost & context

Most cost overruns come from two habits: using a bigger model than the task needs, and letting context bloat. Fix both and you cut spend without touching quality. Run /usage anytime to see where you stand against your limits.

Match the model to the task

• Sonnet handles roughly 80% of work — features, edits, the everyday loop. Keep it as your default.
• Opus for hard multi-file changes, architecture decisions, and tricky bugs that need deep reasoning. Reach for it deliberately, not by default.
• Haiku for trivial, mechanical work where a smaller model is plenty.

Picking Sonnet over Opus for routine work is the single biggest lever. See Models and effort.

Manage context: /clear vs /compact

Context is cost — every token in the window is re-sent on each turn.

• /clear between unrelated tasks. A kitchen-sink session that mixes a bug fix, a refactor, and a doc edit drags all of it into every prompt. Clear and start clean.
• /compact mid-task to summarize a long thread while keeping the thread going.

IMPORTANT

The cache nuance: /compact is cheap only inside the prompt-cache window (about 5 minutes since the last activity). After that the cache has expired, so compacting re-reads the whole conversation — and /clear is the cheaper choice. Rule of thumb: compact while you're actively working, clear when you're switching tasks or coming back cold.

Keep CLAUDE.md lean

A bloated CLAUDE.md is loaded into every session and, worse, causes Claude to ignore your instructions when it's overstuffed. Keep only the few facts that change behavior — build/test commands, where core logic lives, hard conventions. Cut anything Claude can rediscover on its own.

Be specific

Vague prompts are expensive: Claude searches, guesses, and backtracks, and a single fuzzy request can burn 50k+ extra tokens. Name the files, the interfaces, and the done condition up front. Specificity is cheaper than iteration.

TIP

Correcting Claude more than twice on the same task is a cost signal, not just a quality one — the context is now cluttered with wrong turns. /clear and rewrite the prompt instead of pushing through.

Delegate research to subagents

Investigation generates a lot of tokens — file reads, search results, dead ends. Push that work into subagents: each runs in its own context window and returns only a summary, so your main thread stays small and cheap. See The context window for why this matters.

Tune thinking tokens

You can cap extended-thinking budget with the MAX_THINKING_TOKENS environment variable. Lowering it on tasks that don't need deep reasoning reduces spend.

NOTE

Community-reported savings from thinking-token tuning land around 30–40%. Treat that as a guideline to test on your own workload, not a guarantee — measure with /usage before and after. Don't starve genuinely hard tasks of thinking budget just to save tokens.

A quick checklist

1. Default to Sonnet; escalate to Opus only for hard problems.
2. /clear between unrelated tasks; /compact only inside the cache window.
3. Trim CLAUDE.md to the load-bearing facts.
4. Write specific prompts with named files and a done condition.
5. Delegate research to subagents.
6. Test MAX_THINKING_TOKENS tuning on routine work.
7. Check /usage to see usage against your limits.

Why this works

Cost in Claude Code tracks tokens, and tokens track model size and context length. Right-sizing the model and keeping the window tight attack both directly. Anthropic's guidance — be specific, keep CLAUDE.md lean, clear between tasks — is as much a cost strategy as a quality one.

Related: Cheatsheet · Best practices · Learning path

Source: Claude Code best practices; community-reported figures noted as such.