Save Tokens in Claude: 10 Simple Tips That Work

167

To save tokens in Claude, start with how it works: every time you send a message, Claude processes the whole conversation again, not just your newest line. Anthropic’s Claude Code help article says the conversation history is the part that grows fastest, which is why long chats drain usage so quickly. This guide gives you 10 habits, seven that work anywhere in Claude and three for Claude Code, and each one is checked against Anthropic’s own documentation.

I built it from four community guides on Medium, XDA Developers and MyDataSchool, then checked their claims against Anthropic’s help center and docs. Where a popular tip is outdated or contradicts Anthropic, I say so. Numbers reported by individual developers are labelled as self-reported. This is a research synthesis, not a lab benchmark.

Contents

1. Why does Claude use so many tokens?
2. Should you start a new chat for every task?
3. Do batching and specific prompts save tokens?
4. Which Claude model should you use to save tokens?
5. Do effort and thinking settings change token use?
6. Do Projects and clean files save tokens?
7. How do you save tokens in Claude Code?
8. Which token-saving tips are overrated or outdated?
9. Frequently Asked Questions

Why does Claude use so many tokens?

Claude re-reads the whole conversation on every turn, so usage grows with chat length, not just message length. Long threads cost the most.

A token is a small chunk of text, and Claude counts both what you send and what it writes back. Anthropic’s Claude Code help article says each turn sends the conversation so far, your project context and your new prompt, and that the conversation is the item that grows fastest. In the Claude app, Anthropic’s usage guidance also lists attachment size, tools such as Research and web search, model choice and effort level as usage drivers.

Scale explains why this matters. Anthropic’s cost documentation puts average Claude Code spend at about $13 per developer per active day across enterprise deployments, with 90% of users under $30 per active day. “Tokens compound per turn, so manage your context like you manage your money,” says Ian Binder, a chief technology officer, in an April 2026 Medium post.

What adds tokens	Why it costs usage	Fix
Long conversations	The full history is re-read every turn	Tips 1 and 2
Repeated instructions and messages	Each extra turn reprocesses the chat	Tips 3 and 4
Model and effort level	Bigger models and longer thinking use more	Tips 5 and 6
Attachments	File size counts toward limits	Tip 7
Loaded tools and project files	They sit in context every turn	Tips 8, 9 and 10

Should you start a new chat for every task?

Yes. Start a fresh chat whenever the topic changes, because every earlier message keeps being re-read and billed. Carry over only a short summary.

Tip 1: One chat per task. Anthropic’s Claude Code guide names clearing between tasks the first of its five habits, because a session that wandered through three unrelated problems pays for all three on every new message. A simple test: if your next prompt would make sense in a brand-new chat, open one. In Claude Code, /clear does this without closing the session.

Tip 2: Carry a summary, not the history. Ask Claude to list the decisions and outputs you still need, then paste that as the first message of the new chat. In Claude Code, /compact does the same job mid-task, and Anthropic says to use /clear for a new task and /compact to continue a long one. On paid plans, Claude can also search past chats and use memory, so you do not have to re-paste background.

There is no official turn count. Ian Binder suggests starting over before 15 to 20 turns, while MyDataSchool’s Mike Shakhomirov compacts at roughly 30. Both are self-reported rules of thumb. Anthropic’s advice is to switch on topic change, not on a number.

Do batching and specific prompts save tokens?

Yes. One clear, complete message beats three vague ones, because each extra turn makes Claude process the conversation again. Specific prompts also avoid clarifying back-and-forth.

Tip 3: Batch related questions. Anthropic’s usage guidance recommends combining related questions or tasks into a single message and providing background upfront, which cuts the number of turns.

Tip 4: Be specific. The Claude Code documentation contrasts a vague request like “improve this codebase”, which triggers broad scanning, with a precise one that names the function and file. The same logic works in chat: state the audience, format and length in the first message. Our guides to writing better AI prompts and prompt engineering for beginners show how.

Which Claude model should you use to save tokens?

Sonnet 5.5 handles most everyday work at half the per-token price of Opus 5.5. Save Opus for hard, high-stakes tasks that need careful judgment.

Tip 5: Match the model to the job. Anthropic’s launch page lists Sonnet 5.5 at $2 per million input tokens and $10 per million output, against $4 and $20 for Opus 5.5. Anthropic’s help center says Opus uses meaningfully more of your quota, and suggests Sonnet for routine work and Haiku for quick lookups. Anthropic adds that Sonnet 5.5 costs up to 30% less per task than Sonnet 5, and finance-firm Balyasny reports about 121,000 tokens per answer against 497,000 for Sonnet 5 (a customer figure, not independently verified).

Cheaper is not always cheaper. In our Sonnet 5.5 vs Opus 5.5 comparison, Artificial Analysis data shows Sonnet costing more per task than Opus at max effort ($7.60 against $5.98). Anthropic also suggests planning with Opus and executing with Sonnet, which Claude Code supports through /model opusplan.

Do effort and thinking settings change token use?

Yes. Higher effort makes Claude think longer, and thinking tokens are billed as output. Lower the effort setting for simple tasks like edits and summaries.

Tip 6: Lower effort for simple work. Anthropic’s documentation says the default effort is Medium in the Claude apps and Claude Code, and High on the API platform. It also says thinking cannot be turned off on Opus 5.5 or Sonnet 5.5 in Claude Code, so effort is your lever there, set with /effort. One XDA Developers writer reported that dropping from High to Medium cut output tokens by 45% across five tests, a self-reported result I could only confirm from a search summary. Turn off Research and web search when you do not need them, since Anthropic lists both among usage drivers.

Do Projects and clean files save tokens?

Yes. Files stored in a Project are cached, so reusing them counts less against your limits than uploading them again in every chat.

Tip 7: Use Projects for repeated documents, and trim the rest. Anthropic says Project content counts less when reused while the cache is warm, but the cache expires after inactivity, so your first message after a long break counts in full. Upload your core documents once, at the start of a project, and ask many questions against them.

For one-off files, send only what you need. Ian Binder reports seeing 10 to 20 times more tokens from PDFs and webpages than from clean markdown, a self-reported figure well above anything Anthropic publishes, so treat it as an upper-end anecdote. The direction is sound, though: an XDA writer also strips noise and images from long PDFs before sending them, and Anthropic lists attachment size as a usage factor.

How do you save tokens in Claude Code?

Keep CLAUDE.md short, point Claude at exact files, and plan before big changes. Those habits stop Claude Code reading and rewriting more than necessary.

Tip 8: Keep CLAUDE.md under about 200 lines. CLAUDE.md loads at the start of every session, so every line is paid for on every message. Anthropic says to aim for under 200 lines, to move workflow-specific instructions into skills that load on demand, and to add a note only the second time you correct Claude on the same thing. An XDA writer follows the same split: stable rules in CLAUDE.md, one-off instructions in the prompt.

Tip 9: Point at files and plan first. Anything you paste stays in context for the whole session, so name the file and function instead. One correction to a popular tip: Medium’s Habib Mohammed recommends @file references, but Anthropic’s help center notes that @ injects the entire file, so a bare path saves more. Anthropic also says a plan costs a few hundred tokens, while a wrong 400-line diff you revert and regenerate costs thousands.

Tip 10: Audit with /context and trim tools. Run /context to see what is loaded before you type anything. One XDA writer reported about 51,000 tokens in use before the first prompt, a self-reported result that shows why the check is worth two minutes. Disable servers you are not using with /mcp, and prefer command-line tools like gh where they exist.

Which token-saving tips are overrated or outdated?

Some popular tips save little or are outdated. Terse “caveman” prompts, skipping thank-yous, and stripping every MCP server matter far less than clearing long chats.

• MCP overhead is smaller than it was. Habib Mohammed’s April 2026 post says every connected server loads its full tool definitions each session. Anthropic’s current documentation says tool definitions are deferred by default, so only names enter context until a tool is used. Still disable servers you never use.
• Caveman prompts are a trade-off. Shakhomirov shows a shortened prompt using about a quarter of the input tokens and calls the style risky because Claude has to infer more. Anthropic’s advice is specific, not terse, and history costs more than any one prompt.
• Thank-yous cost turns, not words. The words are nearly free, but each extra message makes Claude process the whole chat again, so end long threads rather than closing them politely.

Frequently Asked Questions

What is a token in Claude?

A token is a small chunk of text. Claude counts the text you send and the text it writes back toward your usage. Longer conversations, larger files and longer thinking all increase the total, which is why several habits in this guide focus on keeping context small.

Does Claude Pro have unlimited usage?

No. Anthropic says paid plans have a five-hour session limit plus weekly limits, and both are visible under Settings then Usage. Message length, attachments, conversation length, tools, model choice and effort level all affect how quickly you use them up.

What can I do when I hit my limit?

You can wait for the reset or turn on usage credits. Anthropic says usage credits let Pro and Max subscribers keep working at standard API rates after the plan limit, and you can set a spend limit and alerts under Settings then Usage.

Does a new chat lose my context?

Yes, within the chat itself. Paste a short summary as your first message. Anthropic says paid plans can also search past chats and use memory to carry context forward, so you do not have to re-explain background every time.

Which single habit saves the most?

Starting a fresh chat for each task. Anthropic describes clearing between tasks as the most effective lever for both cost and quality in Claude Code, and the same logic applies in the app: the shorter the history, the less Claude re-reads.

How this guide was checked

Tips were checked against Anthropic’s help center and Claude Code documentation on September 30, 2026. The XDA Developers guide could not be fetched in full, so I used its search summary plus a second XDA article for the figures cited above. Prices and defaults change quickly, so treat them as correct on that date.

[ADAM: add 2 to 3 sentences from your own daily use, for example what changed when you moved to one chat per client task. This is the experience signal Google and AI systems reward. Delete this line if you have nothing real to add.]

Related reading: ChatGPT vs Claude vs Gemini: which is best for beginners | Free AI tools for beginners