How to minimise your Claude token usage without sacrificing quality
In brief
Tokens are what you pay for. The practical ways to use fewer of them, from how you prompt to which model and effort level you choose. Updated for the November 2026 models.
Contents
Short answer: keep instructions short, start new chats for new topics, upload only what Claude needs, ask for shorter answers, and lower the effort setting or use a smaller model for routine work.
Every message you send Claude, every document you upload, every response Claude generates — all of it is measured in tokens. Tokens are how usage is tracked (on Claude.ai plans) and how costs are calculated (on the API). Using fewer tokens does not mean using Claude less. It means using Claude more efficiently.
The biggest token costs (and what to do about each)
1. Long system prompts and Project instructions
Your system prompt or Project instructions are sent with every single message. If your instructions are 3,000 tokens, that is 3,000 tokens consumed on every turn of the conversation — even if the user just says "thanks."
What to do: Keep instructions focused and specific. Remove examples that illustrate the same point. Cut the "do not" lists — Claude follows positive instructions better than negative ones anyway. A tight 500-token system prompt that covers the essentials outperforms a 3,000-token one that covers every edge case.
2. Uploading large documents
When you upload a document to a Project or a conversation, the entire document is loaded into the context window. A 50-page PDF can be 30,000+ tokens.
What to do: Only upload documents Claude actually needs. If Claude only needs your product FAQ, don't also upload your entire employee handbook. For large documents, extract the relevant sections rather than uploading the whole thing. If you reference the same documents repeatedly, consider using prompt caching on the API. In a Project, large knowledge bases switch to retrieval automatically, so Claude only reads the parts it needs.
3. Long conversation histories
Every message in a conversation — yours and Claude's — stays in the context window. A 40-message conversation can consume 20,000+ tokens before you type anything new.
What to do: Start fresh conversations for new topics rather than continuing old ones. In Claude.ai, each new chat in a Project still gets your Project instructions, so you don't lose context. On the API, manage conversation history deliberately — trim or summarise old messages rather than sending the entire history every time.
4. Verbose Claude responses
By default, Claude gives thorough, detailed responses. If you only need a short answer, you're paying for paragraphs you don't read.
What to do: Tell Claude how you want the response formatted. "Answer in 2-3 sentences" or "give me a bulleted list, no explanation" or "just the number." Adding output format instructions to your system prompt ensures every response is appropriately concise.
Choosing the right model
Not every task needs the most powerful model. As of November 2026 the lineup is:
- Haiku 5.5: fastest and cheapest. Good for simple tasks like sorting, extraction and reformatting.
- Sonnet 5.5: the everyday default for drafting, summarising, analysis and most coding.
- Opus 5.5: for complex reasoning and difficult problems.
- Fable 5.2: the most capable model, for the hardest work and long-running agents.
In Claude apps, you choose the model in the model picker for each chat. On the API, sending simple tasks to Haiku 5.5 and harder ones to Sonnet 5.5 or Opus 5.5 can cut costs a lot without hurting the simple tasks. See Which Claude model should your team use? for more.
The lever most teams miss is effort. Every current model has an effort setting: low, medium, high, xhigh and max. It controls how much Claude thinks, and thinking uses tokens. Routine work does not need the setting meant for hard coding problems. Lower effort before you switch to a smaller model. It is the cheaper experiment, and it often makes the switch unnecessary.
Connectors and token awareness
Connectors pull in external content — documents from Google Drive, messages from Slack, pages from Notion. Every piece of content a connector retrieves consumes tokens.
Best practice: Disable connectors you don't need for a given conversation. If you are writing marketing copy, you probably don't need your Jira connector active — it will pull in issue context that adds tokens without adding value. Enable only the connectors relevant to your current task.
The 80/20 rule
For most teams, two changes make the biggest difference:
- Tighter system prompts — cut yours in half, test, iterate
- Start new conversations instead of running one endlessly
Everything else is optimisation. Get these two right first.
Quick answers
What uses the most tokens in Claude? Long instructions that repeat with every message, large uploaded documents, long conversation histories and long replies. Connectors add tokens for everything they pull in.
Does a long conversation use more tokens? Yes. Every earlier message stays in the context window, so a 40-message chat can use 20,000 tokens or more before you type anything new. Start a fresh chat for a new topic.
What is the cheapest Claude model? Haiku 5.5. It suits simple jobs like sorting, extraction and reformatting.
Should I lower effort or switch to a smaller model? Try lowering effort first. It is the cheaper experiment, and it often makes switching models unnecessary.