# How to reduce token usage in Claude and Claude Code > Batch your messages, clear context, trim CLAUDE.md and MCP servers, and match the model and effort to the task, according to Anthropic's docs. URL: https://claudegpt.si/news/reduce-token-usage-in-claude Published: 2026-11-02T15:49:46.581Z Author: ClaudeGPT Si Newsroom Section: Guides Topics: Claude, Claude Code, Anthropic, Token usage, Usage limits Publisher: ClaudeGPT Si ## Key takeaways - In the Claude app, batching questions, reusing Projects and starting new chats for new topics lowers how fast you hit limits. - In Claude Code, /clear between tasks, a CLAUDE.md under 200 lines and disabling unused MCP servers keep context small. - Anthropic says Sonnet costs less than Opus for most coding tasks, and lower effort levels cut thinking tokens. To reduce token usage in Claude, keep conversations short, put everything Claude needs into one message, reuse files through Projects, and lower the effort level or pick a smaller model for easy work. In Claude Code, the biggest levers are clearing context between tasks, trimming CLAUDE.md and unused MCP servers, and matching the model to the job. This guide covers both, from Anthropic's [Claude Help Center](https://support.claude.com/en/articles/9797557-usage-limit-best-practices) and the [Claude Code costs documentation](https://code.claude.com/docs/en/costs), as of November 2, 2026. It is for anyone who keeps hitting a session or weekly limit, or who pays per token through the API. To see where you stand first, read our guide on [how to check Claude token usage](https://claudegpt.si/news/claude-how-to-check-token-usage). ## What uses up your Claude limits Anthropic's help center lists what drives consumption: message length, file attachment size, conversation length, tool use such as Research and web search, the model you pick, the effort level, artifact creation, and multi-step tasks like running code or browsing. Shorter conversations, lighter tool use and lower effort all reduce it. The Claude Code docs add a point that surprises people. Claude Code sends the full conversation with every request, so a one-line question in a session that has been open all day still draws usage for the whole history. Prompt caching makes that cheaper, but not free. ## How to use fewer tokens in the Claude app These tips come from the help center page on usage limit best practices. 1. **Batch your questions.** Combine related questions or similar tasks into one message instead of sending several. Give background up front and be specific, so Claude doesn't need to ask clarifying questions. 2. **Don't retype context.** Within a chat, refer back to earlier information. On paid plans (Pro, Max, Ultra, Team, Enterprise) Claude can also search past chats and use memory and project summaries. 3. **Put repeated documents in a Project.** Anthropic says cached project content counts less against your limits when it is reused. Caches expire after a period of inactivity, so the first message after a long break counts that content in full again. Projects also offer a retrieval (RAG) mode for larger knowledge bases. 4. **Start a new chat for a new topic.** Conversation length is one of the listed drivers, so a long, unrelated thread costs more per message than a fresh one. 5. **Turn down what you don't need.** Skip Research and web search when a plain answer will do, and use a lower effort setting for simple questions. 6. **Watch your meters.** Open Settings > Usage to see progress bars for the five-hour session limit and the weekly limits. For coding, the help center suggests giving the full environment in your first message and sending complete code snippets in one message. For writing, state the requirements, audience and key points up front, and send entire texts for editing at once. ## How to reduce token usage in Claude Code Claude Code bills by API token consumption, or draws from your plan's limits. Anthropic's docs say the average enterprise cost is around $13 per developer per active day, and that 90% of users stay below $30 per active day. These steps keep you toward the low end. ### Clear and compact context - Run `/clear` when you switch to unrelated work. Stale context wastes tokens on every later message. Use `/rename` first so you can find the session again with `/resume`. - Run `/compact` with instructions, for example `/compact Focus on code samples and API usage`, to tell Claude what to keep. You can also set this permanently in CLAUDE.md under a "Compact instructions" heading. - Compacting is itself a large request when the context is big. When you want a fresh start instead of continuity, `/clear` costs nothing. ### Pick the right model and effort Anthropic says Sonnet handles most coding tasks and costs less than Opus, so reserve Opus for complex architecture or multi-step reasoning. Switch with `/model` or set a default in `/config`. For simple subagent work, set `model: haiku` in the subagent configuration. Extended thinking tokens are billed as output tokens. For simple tasks, lower the effort with `/effort` or in `/model`. On some models you can also disable thinking in `/config`, but Anthropic notes you can't turn it off on Opus 5.5, Sonnet 5.5, Haiku 5.5 or the Fable models. On models with a fixed thinking budget, `MAX_THINKING_TOKENS=8000` caps it. Adaptive-reasoning models ignore nonzero budgets, so use effort levels there. ### Trim what loads at the start - **CLAUDE.md:** it loads at session start, so detailed workflow instructions cost tokens even when you're doing something else. Anthropic suggests keeping it under 200 lines and moving specialized instructions into skills, which load only when invoked. - **MCP servers:** run `/context` to see what is using space, and `/mcp` to disable servers you aren't using. Tool definitions are deferred by default, but command-line tools such as `gh`, `aws` and `gcloud` are still more context-efficient than MCP servers. - **Hooks:** a hook can filter output before Claude sees it, for instance returning only the `ERROR` lines of a 10,000-line log. The docs include a PreToolUse hook that shows only test failures. ### Work in a way that avoids rework - Write specific prompts. "Add input validation to the login function in auth.ts" triggers far fewer file reads than "improve this codebase". - Use plan mode (Shift+Tab) for complex tasks so Claude proposes an approach before editing. - Press Escape as soon as Claude heads the wrong way, and use `/rewind` to restore a checkpoint. - Delegate noisy jobs such as running tests or reading logs to subagents, so only a summary returns to your main conversation. Their requests still draw on your usage, so give them a smaller model. - Keep agent teams small. Anthropic says they use roughly 7 times more tokens than standard sessions when teammates run in plan mode. ### Let `/usage` tell you what to change On Pro, Max, Ultra, Team and Enterprise plans, `/usage` shows what counted against your limits, split by skills, subagents, plugins and MCP servers, plus recommendations for anything behind 10% or more of recent usage. Press `f` to have Claude Code apply the ones it can (this needs Claude Code v2.1.305 or later). ## Hidden sources of usage Anthropic's docs list several things that spend tokens while you aren't typing: - The first message after a break longer than the cache lifetime reprocesses your whole context. - Session-scoped scheduled tasks, such as `/loop`, fire on their interval even when the session is idle. - Session recap and idle compaction can use tokens while you're away, and prompt suggestions send a short request after each response. All three can be turned off. - Messages from another of your sessions, and goal check-ins, can start new turns in an idle session. If you hit a limit anyway, `/usage-credits` opens the settings for usage credits, which let you continue past your plan's allowance. ## FAQ ### How do I check how many tokens I've used in Claude? In the Claude app, open Settings > Usage for the session and weekly bars. In Claude Code, run `/usage`. Our [token usage guide](https://claudegpt.si/news/claude-how-to-check-token-usage) has the steps. ### Does starting a new chat save tokens? Yes, in most cases. Conversation length counts toward limits, and Claude Code resends the full history with each request. Run `/clear` in Claude Code or start a new chat in the app when the topic changes. ### Does a shorter prompt use fewer tokens? Message length does count, but specific prompts matter more. A precise request avoids broad searching, extra file reads and clarifying back-and-forth. ### Which Claude model uses the least? Anthropic says Sonnet costs less than Opus and suits most coding work, and recommends Haiku for simple subagent tasks. Lower effort levels also reduce usage. Check the usage page again after a session of trying these changes. The `/usage` recommendations will show whether context or cache misses are still the main cost. ## Frequently asked questions ### How do I check how many tokens I've used in Claude? In the Claude app, open Settings > Usage to see the five-hour session and weekly bars. In Claude Code, run /usage for session token totals and, on subscription plans, a breakdown with recommendations. ### Does starting a new chat save tokens? Usually. Conversation length counts toward limits, and Claude Code resends the full history with every request. Run /clear in Claude Code or open a new chat in the app when the topic changes. ### Does a shorter prompt use fewer tokens? Message length counts, but specific prompts matter more. A precise request avoids broad file scanning and clarifying back-and-forth. ### Which Claude model uses the least? Anthropic says Sonnet costs less than Opus and handles most coding tasks, and suggests Haiku for simple subagent work. Lower effort levels also reduce usage. ## Sources - [Manage costs effectively](https://code.claude.com/docs/en/costs), Anthropic - [Usage limit best practices](https://support.claude.com/en/articles/9797557-usage-limit-best-practices), Anthropic