How to save Claude tokens

8 concrete strategies that let you use up to 5x fewer tokens — without sacrificing quality.

1. Your entire history is re-read with every message

Claude does not remember the conversation. It reads everything again from the start, every time you send a message. At message 1 that is 1 block of history, at message 5 it is 5, at message 10 already 10.

2. Token cost vs token burn

Two different things, and you can influence both.

Cost is the multiplier — it depends on which model you choose: Haiku 1x, Sonnet 3x, Opus 5x.

Burn is the volume — it depends on your habits:

• Using PDF instead of Markdown files
• 500 words of output when you need 1 sentence
• 30 messages when you should have started over at 10

Reduce either one, and you will notice it immediately.

3. Going back and forth

The biggest token waster for most users. Do not send change requests one by one — send a single prompt with everything in it.

Expensive: "Make it bigger", "Actually a bit smaller", "A bit more towards the center", "Wrong color, a bit redder", "Different text".

Efficient: "Padding 64 to 80px, heading 48px semibold, center the CTA button, use #2216FF as the primary color."

Same result, but with 5x fewer tokens used.

4. The MCP tool tax

Tools are reloaded with every message, even if you do not use them.

• No tools: 0 tokens per message (0% of Pro)
• Figma MCP: ~4,000 tokens per message (9% of Pro)
• 4 tools (Notion, Figma…): ~7,000 tokens per message (16% of Pro)
• 5+ tools: ~55,000 tokens per message (125% of Pro)

Solution: turn tools off with /mcp in Claude Code, or disconnect them in Claude.ai via Settings → Tools.

5. Using the wrong model

Opus is more powerful but ~5x more expensive than Sonnet for many tasks.

• Rewriting text → Sonnet
• Simple analysis → Sonnet
• Brainstorming → Sonnet
• Summary → Sonnet
• Complex automations → Opus

98% of tasks work fine on Sonnet. That is a 5x saving compared to Opus for the same result.

6. Limiting the output

Output tokens cost 5x as much as input tokens. Tell Claude what you need.

Claude's default: an 800-word essay with context, explanations, suggestions, follow-up questions and caveats you never asked for (~1,200 tokens). What you actually need: 3 problem points regarding the forms, each in one sentence (~80 tokens).

Add these prompts:

• "Just the code, no commentary"
• "Answer in one sentence"
• "3 bullet points max"
• "Give me the diff only"
• "No explanations"
• "Just list the things"

7. Which model for which task?

A quick reference. Choose the lightest model that can handle the task and save tokens immediately.

• A translation → Haiku
• A task that needs to be fast → Haiku
• Brainstorm → Sonnet
• Rewriting text → Sonnet
• Summary → Sonnet
• Simple analysis → Sonnet
• Complex automations → Opus

8. When should you start a new conversation?

Knowing when to reset is just as important as being efficient. Reset triggers: the task is done, the topic changes, 15–20 messages, Claude gets confused.

• /compact — Summarize & shorten. Claude condenses your conversation history into a shorter version. Useful when you need context but want to save tokens.
• /clear — Full reset. Wipes the entire history. Use this when you start a new task or the context is no longer relevant.
• New chat — Simply start a new conversation in claude.ai. Your project data is kept, the chat history is not.