Claude Code: 7 Strategies to Control Token Costs

Claude Code: Understanding and Optimizing Token Usage

Claude Code, a powerful tool for developers, can quickly become a source of significant expenses if its usage is not optimized. Indeed, beyond the simple prompt you enter, Claude often retains the entire session, including previous messages, already analyzed files, tool outputs, and other background instructions. Thus, when token usage increases, the problem usually stems from an overloaded context rather than an ineffective prompt.

Generic advice such as "keep conversations short" is not enough to solve the problem. What truly makes a difference is understanding how Claude Code constructs and utilizes context, what is constantly returned, and which parts of your workflow subtly add waste. This article presents 7 practical methods for using Claude Code effectively while managing costs.

1. Match the Model to Task Complexity

A simple yet often overlooked method is to choose the appropriate model based on the task. Not all tasks require the use of the most expensive model. For example, in API billing, the Opus model costs five times more per token than the Sonnet model. In subscriptions, heavier models consume your quota more quickly.

• /model sonnet: Ideal for everyday tasks such as writing tests, simple modifications, and code explanations.
• /model opus: Recommended for multi-file architectural decisions and debugging complex issues.
• /model haiku: Suitable for repetitive tasks like searching, formatting, and renaming.

Start each session with Sonnet and only switch to Opus when you truly need in-depth analysis or complex refactoring. Haiku is perfect for mechanical tasks. You can also adjust the effort level with /effort to save tokens on simple tasks.

2. Optimize the Use of CLAUDE.md

To save tokens, avoid repeating the same project rules in every conversation. CLAUDE.md is designed for this purpose. It loads before Claude reads your code or task and persists in the context window throughout the session. A 5,000-token CLAUDE.md costs 5,000 tokens each turn, whether you send 2 messages or 200. Place your stable instructions there: how to run tests, which package manager to use, your formatting rules, and important architectural constraints.

Make sure to keep CLAUDE.md concise. Avoid including meeting notes, design histories, or lengthy implementation guides. You will achieve better results when CLAUDE.md functions as a reference table rather than a dumping ground for ideas.

3. Use Sub-Agents for Verbose Tasks

Utilizing sub-agents is an effective strategy for managing context. Sub-agents are isolated instances of Claude that operate in their own context window. When a sub-agent runs, all its verbose output remains isolated, and only the summary is returned to your main conversation. This helps keep your main thread cleaner.

However, sub-agents are not automatically cheaper. For simple tasks, such as shell actions or quick git operations, a sub-agent can be costly due to the overhead it adds. The rule is to use sub-agents when reducing clutter in the main context is worth more than the startup overhead.

4. Precisely Target Files and Lines

Asking Claude to "look in the repository" without precision can lead to token waste. The vaguer the task, the more likely Claude is to open multiple files unnecessarily. For example:

• "Look at the authentication code and tell me what's wrong."
• "Compare lines 30 to 90 of src/auth/session.ts with lines 10 to 60 of src/api/login.ts and explain the inconsistency."

The first request seems natural, but it can trigger costly exploration. Use plan mode before costly operations by activating Shift+Tab. This allows Claude to produce a step-by-step plan without modifications, which you can then refine before returning to normal mode.

5. Proactively Use /compact

Claude can automatically compact your session, but you can also execute /compact yourself. Timing is crucial. After Claude has inspected several files and executed commands, your session may contain a lot of unnecessary material. This is the ideal time to compact.

A common mistake is to wait until Claude starts forgetting information or shows a context warning before compacting. At this point, the session is already overloaded, and the summary is not as clear. By compacting earlier, when the session is still "healthy," you retain key information and eliminate noise.

6. Check /context Before Optimizing

Before modifying your workflow, examine what is actually consuming the context. Much token waste seems mysterious until you realize that the costly part may be a large file that Claude read earlier or accumulated tool output.

The /context command is your diagnostic tool. Before changing your entire workflow, identify what is loaded or returned multiple times. Often, the biggest improvement comes from spotting a "silent offender" present in every turn.

7. Simplify Your Tool Configuration

Claude Code can connect to many external tools, but more tools also mean more context overhead. If too many tools are involved, the model may carry more overhead than necessary. Keep your setup lightweight and use integrations that solve a real recurring problem.

Conclusion

The key to reducing Claude Code's token usage is not to monitor every prompt but to design your workflow so that Claude only sees what it needs. The greatest gains come from controlling automatic context, narrowing the search scope, and preventing noisy side work that contaminates the main session.