7 Practical Ways to Reduce Claude Code Token Costs

Claude Code costs rise when each turn carries a large conversation history, broad file context, verbose tool output, and instructions that are loaded repeatedly. The most effective savings come from controlling what enters the context, not from making every prompt unnaturally short.

1. Match the model to the task

Use the model selector available in your Claude Code version and account. Reserve the most capable—and usually more expensive—model for architecture decisions, difficult debugging, and work that genuinely needs deeper reasoning. Routine edits, explanations, formatting, and focused test changes often need less capacity.

Model names, availability, and prices change, so check the current model list and Anthropic pricing rather than relying on a fixed cost comparison.

2. Keep CLAUDE.md focused

CLAUDE.md is valuable for stable project guidance: build and test commands, package-manager rules, formatting conventions, architectural boundaries, and directories that must not be changed. Remove meeting notes, old plans, duplicate explanations, and instructions that apply to only one task.

A short, specific file is easier for the model to follow and avoids repeatedly carrying irrelevant context. Anthropic's guidance also recommends keeping project instructions concise and focused. For a related workflow, see how to help Claude Code retain useful project lessons.

3. Give Claude a narrow starting point

Instead of asking Claude to inspect an entire repository, name the symptom, relevant files, expected behavior, and the command that reproduces the problem. For example:

Fix the failing validation in src/forms/account.ts.
Reproduce it with: npm test -- account-form
Keep the public API unchanged and show the test result.

This still leaves room for investigation, but it prevents an unnecessary first pass across unrelated directories.

4. Control large tool output

Logs, test reports, generated files, and MCP responses can consume more context than the prompt itself. Filter commands to the failing test, relevant time range, or matching error. Prefer a concise summary plus the important lines over pasting thousands of lines.

Also review enabled MCP servers. Keep integrations that solve recurring tasks, and disable ones that add tool descriptions or output you rarely use. Anthropic documents configurable limits for large MCP results in its Claude Code MCP guide.

5. Use subagents for genuinely separate work

A subagent can isolate a large search, log analysis, or research task and return a compact result to the main conversation. That can protect the main context during complex work. It also adds its own instructions and calls, so it is not automatically cheaper for a one-command task.

Use subagents when the isolated work would otherwise flood the main session, such as exploring several independent modules. Keep quick file lookups and simple commands in the main session.

6. Compact at a useful checkpoint

After investigation is complete—and before a long implementation phase—use Claude Code's current context-management controls to summarize the session. A good checkpoint is when the cause, constraints, affected files, and next steps are known but the conversation contains discarded hypotheses or lengthy output.

Before compacting, state the facts that must survive: the confirmed cause, decisions, files changed, tests to run, and unresolved risks. If a session has drifted across unrelated tasks, starting a new focused session may be clearer than compressing it.

7. Measure before optimizing

Use the context or usage information available in your Claude Code version to identify the largest contributors. Common causes include a long instruction file, a broad repository scan, repeated full test output, or an integration returning oversized responses.

Change one source of overhead, then compare usage on a similar task. For non-interactive automation, structured output can make cost and turn information easier to record. Anthropic's CLI reference lists the currently supported flags.

A cost-conscious workflow

1. Start a session for one defined outcome.
2. Point to the likely files and a reproducible command.
3. Use the least expensive model that can complete the task reliably.
4. Filter logs and tool results before adding them to context.
5. Compact at a phase boundary, or start a clean session for a new task.
6. Review the diff and run focused tests before widening the scope.

Cost reduction should not come at the expense of verification. A narrowly scoped, well-tested change is usually both cheaper and safer. See also how to use Claude Code safely when deciding which tools and permissions to allow.