Table of Contents
Overview
Using Claude Code day to day, token consumption can pile up before you realize it. Here's a prioritized rundown of concrete ways to hold down token usage.
Don't load an entire large file at once
The most basic fix is not feeding an entire large file to the AI as-is. Passing only the part you actually need, or loading it in chunks, prevents wasted token consumption. Handing over the whole file might feel convenient, but the cost climbs steadily as a result.
Don't ask for unnecessarily verbose output
Another thing that's easy to overlook is the sheer volume of the AI's own output. Asking for overly polite explanations or a large volume of comments increases the token count of the output itself. Instructing the AI at the prompt stage to output only what's truly necessary, concisely, can also help cut costs.
Reset your session frequently
The longer a conversation runs, the more information piles up in context. Making a habit of frequently resetting the session with the "/clear" command keeps you from dragging along an unnecessarily bloated context. Setting a rule to reset once a task is finished seems like a workable habit.
Switch models depending on the task
For minor edits or simple checks, switching to a lower-tier model like Haiku can cut costs substantially. Not every task needs to be run through the top-tier model — matching the model to how heavy the task is seems to lead to more efficient operation.
Understand prompt caching
If you're using Claude via the API, understanding "prompt caching" can help a lot with cost reduction. This mechanism stores information you've sent once on the server side for a certain period, so subsequent instructions can reuse the same information. The cost of reusing cached information is reportedly around 10% of loading it fresh — a meaningful cost reduction when conditions line up.
The trick to making good use of this is ordering your prompt with "information that doesn't change" first and "information that does change" later. Concretely, putting content that's the same every time — system settings, tool definitions — in the first half, and content that changes, like conversation history, in the second half, makes the cache more likely to kick in. The cache is typically held for around 5 minutes, but a relatively recent version of Claude Code reportedly lets you extend that retention to up to an hour. If you're working within the same project context over a long stretch, it's worth checking this setting.
Revisiting the highest-priority countermeasures
Among everything covered so far, these three are said to be especially effective:
- Use Plan Mode to head off unnecessary code generation
- Cut off long conversations as needed and switch to a fresh session
- Maintain CLAUDE.md to reduce the cost of explaining things every time
These aren't just about cutting token usage — they're also said to double as effective ways to avoid the 5-hour rate limit.
Make a habit of tracking usage
Before putting any of this into practice, it matters to make a habit of regularly checking usage with the "/usage" command. Rather than managing cost purely on gut feeling, keeping an eye on the actual numbers — "how much have I used this week" — makes it easier to notice waste.
Summary
Practical techniques for cutting Claude Code costs can be summarized as follows:
- Don't load an entire large file at once
- Don't ask for verbose output
- Reset the session frequently with "/clear"
- Switch to a lower-tier model for light tasks
- Regularly visualize usage with "/usage"
Small everyday habits like these add up to a meaningful difference in your monthly cost. Start with whichever of these you can put into practice today.