How to Reduce Claude Token Usage (Claude Code, Chat and API)

The biggest lever on Claude token usage is context: Claude Code sends your whole conversation with every request, so the fastest savings come from keeping that conversation short. In order of impact: /clear between unrelated tasks, pick the smallest model and effort that does the job, keep CLAUDE.md lean, and push noisy work into subagents. On the API, add prompt caching and the Batch API.

Everything here comes from Anthropic’s Claude Code costs guide, the Help Center’s usage limit best practices and the API pricing page, checked October 25, 2026 unless a row says otherwise. Claudekit.ai is an independent community site, not affiliated with Anthropic.

Start by measuring

You can’t cut what you can’t see. On a Pro, Max, Team or Enterprise plan:

1. Run /usage in Claude Code. Below the plan bars it shows what consumed your allowance (skills, subagents, plugins, individual MCP servers) and recommendations such as “Trim CLAUDE.md · 49k tokens per request”. On Claude Code v2.1.305 or later, press f to have it apply the ones it can.
2. Run /context to see what fills the current context window.
3. On Pro and Max, put your limits on screen with a status line that shows rate limit usage, so you notice a fast climb while it happens.

For every way to read the numbers, including headless runs and team dashboards, see how to check Claude Code token usage.

The rest of this guide is the list of fixes those screens point to.

Claude Code: the levers that matter

Lever	What to do	Why it works
Clear between tasks	/clear when you switch to unrelated work. /rename first so you can /resume later	Stale context is re-sent with every message. /clear costs nothing
Compact with a focus	/compact Focus on code changes and test output	Summarises the conversation, keeping what you name. It reads the whole thread to do it, so it is a large request itself
Right model	Sonnet for most coding, Opus for hard design. model: haiku for simple subagents	Smaller models use less. Switching to Opus also switches subagents that inherit your model
Lower effort	/effort down for simple tasks	Thinking is billed as output and can run to tens of thousands of tokens per request
Lean CLAUDE.md	Keep it under about 200 lines. Move procedures into skills	CLAUDE.md loads at session start; skills load only when used
Subagents for noisy work	Delegate test runs, log digging and doc fetching to subagents	The verbose output stays in the subagent; only a summary returns. Its requests still count, so give it a small model
Fewer MCP servers	Disable unused servers in /mcp. Prefer CLIs like gh where they exist	CLIs add no per-tool listing to the context
Filter with hooks	A PreToolUse hook that trims test output to failures	Hundreds of tokens instead of tens of thousands
Plan first	Plan mode (Shift+Tab) for big changes; Esc to stop a wrong turn early	Re-work after a wrong direction is the most expensive tokens you spend
Specific prompts	“Add input validation to login() in auth.ts”, not “improve this codebase”	Vague requests trigger broad file scanning

Our Claude Code context window guide goes deeper on /context, /compact and what survives compaction, and the CLAUDE.md guide shows how to keep the file short.

Hidden usage in long sessions

Anthropic’s costs guide lists what makes a session that has been open for hours use more than your activity suggests:

• Cache misses after a break. Your first message after the cache lifetime has passed re-processes the full context. After a long break, a fresh session is often cheaper than picking up a huge one.
• Scheduled tasks. A /loop or other session-scoped scheduled task fires on its interval, sending your full context each time, even while you are idle.
• Agent teams. Each teammate runs its own context window. The docs put agent teams at roughly 7x the tokens of a standard session when teammates run in plan mode. Keep teams small and shut teammates down when they finish.
• Background work. Summaries for --resume and similar jobs use a small amount, typically under $0.04 per session per the docs.

Claude chat: make the allowance go further

Based on the Help Center’s tips for claude.ai, the desktop and mobile apps:

• Keep conversations focused. Conversation length is one of the factors that uses your allowance faster, so start a new chat when the topic changes.
• Refer back instead of repeating. On paid plans, chat search and memory let Claude reuse earlier information you would otherwise paste again.
• Batch related questions into one message instead of a string of follow-ups.
• Use Projects for documents you reference repeatedly: cached project content counts less against your limits than new content.
• Be specific so you need fewer follow-up messages.
• Check Settings > Usage to see your session and weekly bars before a big task.

Chat and Claude Code share one allowance, so a long research chat leaves less for coding. The rules are in Claude usage limits explained.

Claude API: pay less per token

On the API you pay for every token, so the same habits save money directly. Two more levers apply:

• Prompt caching. Mark the repeated prefix (system prompt, tool definitions, documents) as cacheable. A cache read costs 0.1x the input price on most current models, and less on some; a cache write costs 1.25x input for 5 minutes or 2x for 1 hour (pricing page, checked October 24, 2026). It pays off when the prefix is reused within its lifetime.
• Batch API. Work that can wait gets 50% off input and output, and caching discounts stack with it.

Run your own numbers in the Claude API cost calculator, and see worked examples in the Claude API pricing guide.

When saving tokens isn’t enough

If you trim everything and still hit limits every week, the issue is capacity, not waste. Compare your usage priced at API rates with Pro, Max and Team in the plan picker, and check the limits changelog for recent changes before you decide.