Claude Code token usage tips: the ones that actually move the number

Context first, routing second, and a two-week comparison so you know which one worked.

·2 min read

Claude Code token usage tips mostly boil down to one sentence: every request resends the conversation, so the cheapest token is the one you removed from context. The UsageAtlas guide covers the same ground. Everything else is a detail of how to remove them without losing the thread you need.

The context rule

Each turn sends the full history. That has two consequences people feel before they understand.

Later turns cost more than earlier ones, even when the edit is identical, because the context behind them is larger.

Idle breaks are expensive. After a pause long enough for the prompt cache to expire, the next request pays full price for context it had already processed.

That is why a quiet afternoon can cost more than a busy one: the activity was small and the context was not.

The changes, in order of return

Change Effort Typical effect
/clear between tasks None Largest single saving
Model per task A habit Large, especially for boilerplate
Lower /effort on simple work A setting Moderate
Shorter CLAUDE.md An edit Moderate, every session
Disable unused MCP servers A setting Moderate, every session
/compact with a note One request Useful when continuity matters

The ordering is deliberate: do the free things before the fiddly ones.

Why compaction is not a free lunch

/compact reads the conversation it summarises, so it is itself a large request. It is worth it when you need continuity and cheaper than letting a thread grow indefinitely, but a clean break with /clear beats it whenever the break is acceptable.

The multipliers

Subagents keep their own context and make their own requests. Agent teams go further: Anthropic's documentation puts them at about seven times the tokens of a single session. Both are appropriate for parallel work and expensive as a default.

Making it stick

Habits decay, so attach the change to a moment. Clear at the start of a new ticket. Switch model when the task changes shape. Check the totals at the end of the week.

The check is the part that keeps the habit honest. UsageAtlas reports Claude Code tokens per day and per model, and a two-week comparison is enough to see whether the change did anything.

What to ignore

Advice that starts with a plan upgrade. A larger allowance buys time, not efficiency, and the same pattern reappears at the new ceiling.

Advice that switches models after a limit. Session and weekly caps cover all models, so the switch must come before the stop.

FAQ

Do prompt caches make long sessions cheap?

They help, and they expire. A break long enough to clear the cache turns a cheap turn into a full-price one.

Is Haiku good enough for subagents?

For file searches and log scans, usually yes. The subagent's job is to return a summary, not to reason about architecture.

Can I see the effect without a tracker?

You can see whether you hit limits less often. You cannot see the size of the change, which is what tells you whether to keep going.

The takeaway

Remove context before you change plans. Clear, route, trim, and measure for two weeks.