The biggest savings in Claude Code come from keeping context small and matching the model to the task. Anthropic's documentation recommends clearing between tasks, choosing a smaller model for simple work, trimming tool servers, lowering reasoning effort when depth is not needed, and writing specific prompts.
The checklist
Everything here comes from Anthropic's Claude Code cost management page, read on 19 October 2026. Check the page for current commands, since they change between versions.
- Clear between unrelated tasks. Stale context wastes tokens on every later message. Rename the session first if you want to resume it.
- Compact with instructions. Tell the summary what to keep, such as test output and code changes.
- Choose the right model. The page says Sonnet handles most coding tasks and costs less than Opus, and suggests reserving Opus for complex architecture or multi-step reasoning.
- Use a smaller model for simple subagents.
- Disable unused tool servers. Check what is consuming context with the context command.
- Prefer command line tools where they exist. The page says tools like gh and aws are more context efficient than servers because they add no per tool listing.
- Lower the reasoning effort on simple tasks. Thinking tokens are billed as output. The page notes some models always use extended thinking.
- Keep your project instructions file short. The page suggests under 200 lines, and moving specialized workflows into skills that load on demand.
- Delegate noisy work to subagents, so long test or log output stays out of the main conversation.
- Write specific prompts. "Add input validation to the login function in auth.ts" beats "improve this codebase".
Work habits that avoid rework
- Use plan mode for complex tasks so the direction is agreed before edits.
- Stop early when the agent heads the wrong way, and rewind to a checkpoint.
- Give a way to verify: tests, expected output or a screenshot.
- Test in small steps.
Preprocess before the model sees it
The documentation describes a hook that filters test output to show only failures, reducing tens of thousands of tokens to hundreds. The same idea works for logs, build output and long diffs: let a script do the searching and show the model the result.
Check, then change
Use the usage screen to see what consumed your allowance. On plans with limits, it flags behaviors such as long context or cache misses when they make up 10 percent or more of recent usage. Change the habit that the screen flags, not the one you suspect.
What not to chase
Background activity such as conversation summaries is small. Anthropic says it is typically under $0.04 per session. Do not spend an afternoon on it.
For the money side, see how Claude plan usage limits work and why agent bills balloon.
Quick check
Does a shorter prompt always save tokens?
Not if it causes more rework. A specific prompt that avoids broad file scanning is often shorter in effect, because fewer files are read.
Should I turn thinking off?
Only for simple tasks and only where the model allows it. Lowering effort is the safer first step.
Sources
Last reviewed: 19 October 2026.
