10 Claude Code Usage Tips to Save Tokens (2026)
Claude Code usage tips are the difference between hitting your session limit by noon and getting a full day of productive work done.
Why Token Management Matters in Claude Code
Most people think Claude Code counts messages. It doesn't. It counts tokens. Every time you send a message, Claude rereads the entire conversation from the beginning. Message one might cost 500 tokens. Message 30 costs 15,000 because it's reprocessing everything before it.
One developer tracked a 100-plus message chat and found that 98.5% of all tokens were spent rereading old history. Only 1.5% went toward the actual output. That means your cost isn't adding up linearly. It's compounding.
On top of your own messages, Claude reloads your CLAUDE.md, MCP tool names and instructions, system prompts, skills, and files on every single turn. This invisible overhead drips into your context constantly. Bloated context doesn't just cost more. It produces worse output thanks to a phenomenon called "lost in the middle," where models pay the most attention to the beginning and end of a session while ignoring what's in between.
If you're new to Claude Code, understanding this token math early will save you serious frustration.
10 Claude Code Usage Tips That Actually Save Tokens
I pulled these from my own experience, community threads, and hours of testing. Organized from simple habits to advanced strategies.
1. Start Fresh Conversations Between Unrelated Tasks
This is the single biggest thing you can do. Use /clear between unrelated tasks. Don't carry context about topic A into a conversation about topic B. Every message in a long chat is exponentially more expensive than the same message in a fresh chat.
At roughly 500 tokens per exchange, 5 messages costs about 7,500 cumulative tokens. By message 30, you're at 232,000. That 30th message alone costs 31 times more than the first one. Fresh chats reset this compounding entirely.
2. Batch Your Prompts Into a Single Message
Three separate messages cost three times what one combined message costs. Instead of sending "Summarize this article," then "Now list the main points," then "Suggest a headline," combine them all into one prompt.
You save tokens twice: fewer context reloads, and you stay further from your session limit. The answers often turn out better too because Claude sees the full picture from the start.
3. Be Surgical With File References
Don't say "here's my whole repo, go find the bug." Say "check the verifyUser function inside auth.js." Use @filename to point at specific files instead of letting Claude explore freely.
If the bug is in one function, paste just that function. Claude needs precision, but so does your input. Every unnecessary file you feed it is tokens burned for nothing.
4. Disconnect Unused MCP Servers
Connected MCP servers used to load every tool definition into your context on every message, and one server alone could add 18,000 tokens. Claude Code now defers those definitions by default with a feature called tool search. Only tool names and server instructions load when a session starts, and a full definition loads when Claude actually needs that tool.
That makes an idle server much cheaper, but not free. The names and instructions still load every session, and tool search switches off in some setups, such as when ANTHROPIC_BASE_URL points to a third-party proxy. There, every definition loads up front again.
Run /mcp at the start of each session and disconnect the ones you don't need. I keep only two or three connected in any given week, and I covered which servers earn their place in a marketing setup in a separate post. Better yet, if a CLI exists for the same tool (like Google Workspace CLI instead of the Google Calendar MCP), use that instead. CLIs are faster and cheaper. The MCP protocol documentation has a full list of available servers and their tool definitions.
5. Use Plan Mode Before Jumping Into Real Tasks
Plan mode lets Claude map out its approach and ask clarifying questions before writing any code. This prevents the biggest source of token waste: Claude going down the wrong path, writing hundreds of lines, and then scrapping everything.
Add something like this to your CLAUDE.md: "Do not make any changes until you have 95% confidence in what you need to build. Ask me follow-up questions until you reach that confidence level."
6. Keep Your CLAUDE.md Lean
Claude auto-reads your CLAUDE.md at the start of every single chat as system context. If your file is 1,000 lines, every message you send triggers a full reload of all 1,000 lines. Even if you just say "hi."
Keep it under 200 lines. Include your tech stack, coding conventions, build commands, and the most important rules. Treat it like an index that routes to where more data lives, not an encyclopedia. Point to skills and reference docs by file path so Claude grabs them only when needed. If you are still deciding which ones to add, start with a curated set of skills worth installing.
7. Compact at 60% Capacity, Not at the Limit
Auto-compact waits until the window is nearly full, about 967,000 tokens (roughly 97%) on a model with a 1 million token window. By that point, your context is already degraded and output quality has dropped. Run /context to check your capacity percentage, and at about 60%, run /compact with specific instructions on what to preserve.
After three to four compacts in a row, quality starts to degrade further. At that point, get a session summary, run /clear, paste the summary into a new chat, and keep going. The full list of available commands is in the Claude Code CLI reference.
8. Long Breaks Cost You (Prompt Cache Timeout)
Claude Code uses prompt caching to avoid reprocessing unchanged context. That cache expires. On a Claude subscription, within your plan's included usage, the main conversation gets a one-hour cache. On an API key or a cloud provider it is five minutes by default, and on a subscription it drops to five minutes once you go past your plan limit and start drawing on usage credits.
Step away for longer than your cache lasts and your next message reprocesses the whole conversation without the cache discount, which burns noticeably more of your usage. This is why some people feel their usage randomly spikes after coming back from lunch. If you're stepping away for a while, do a /compact or /clear before you leave, so that first message back is not re-reading a giant context.
9. Pick the Right Model for the Task
Choosing the right model is the most impactful decision you make daily. Opus 5.5 is now the Claude Code default on every paid plan, and it is what I run for most of my own work. Sonnet 5.5 costs half as much per token and handles most coding and bulk work well. Haiku 4.5 costs half of Sonnet and handles formatting, brainstorming, simple tasks, and sub-agents perfectly.
The second lever is effort. You can't turn thinking off on Opus 5.5 or Sonnet 5.5, so /effort is the control. Drop it to low for mechanical work and save the higher levels for the hard problems.
You can also bring in external tools like Codex for code reviews. Let Sonnet and Haiku handle the volume work, like batch jobs and sub-agents, and keep Opus for the tasks where judgment matters. If you're running agent teams, delegate research and one-off tasks to Haiku sub-agents whenever possible. Check the Claude pricing page to compare costs across tiers.
10. Stop Scheduling Around Peak Hours
This tip used to tell you to push heavy work into the evening. That advice is now dead, and it is worth knowing why.
Anthropic briefly drained your 5-hour session window faster during weekday peak hours. On May 6, 2026 it removed the peak hours limit reduction on Claude Code for Pro and Max accounts, and doubled the five-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans at the same time.
So on Pro and Max there is no longer a clock to work around. Run the big refactor at 9 AM if that is when you are sharp. The nine tips above are what actually move your usage now, and they matter more than the hour on your calendar.
What the Claude Code Leak Revealed (and What Shipped)
Anthropic's source code leak earlier this year (512,000 lines of code across nearly 2,000 files) revealed several features that were unreleased at the time. Here is where they stand now.
Kairos showed up in the leak as an always-on background mode with an "auto-dream" feature, modeled on how humans consolidate memories during sleep. As described, it would review your chat history while you were away, summarize what happened, and store compressed context in memory files. Anthropic has not documented it as a Claude Code feature.
Ultra Plan sent complex planning to a cloud session that you reviewed in your browser. It did ship, as /ultraplan in April 2026, and Anthropic has since removed it. The Claude Code docs now point you to plan mode instead, which is tip 5 above.
Mythos was the leak's model tier above Opus, and it has since shipped. Claude Mythos 5 and Mythos 5.1 are now on Anthropic's pricing page at $10 per million input tokens and $50 per million output, the same rate as Fable 5.1 and two and a half times what Opus 5.5 costs. Access is by invitation only, through a program called Project Glasswing.
So of the three, only Mythos is something you can use today, and only by invitation. The ten tips above are what actually move your usage.
Ryan's Final Thoughts
Token management isn't about being cheap. It's about being intentional. The people who master these habits don't just save money. They get better output, longer sessions, and fewer interruptions. Start with fresh chats and batched prompts. Add the advanced strategies as they become second nature. Your future self will thank you.
Claude Code Usage Tips FAQs
How many tokens does Claude Code use per session?
It depends on conversation length, MCP servers, and loaded files. A fresh session with no chats already uses tens of thousands of tokens from system prompts, tools, and configuration files, and /context shows you the breakdown. A 30-message conversation can easily hit 232,000 cumulative tokens due to compounding context reloads.
Does plan mode actually save tokens in Claude Code?
Yes. Plan mode prevents Claude from writing code down the wrong path, which is the single biggest source of wasted tokens. By clarifying the approach upfront, you avoid costly redo cycles where Claude generates hundreds of lines that get scrapped entirely.
What is the prompt cache timeout in Claude Code?
It depends on how you pay. On a Claude subscription, within your plan's included usage, the main conversation gets a one-hour cache. On an API key or a cloud provider it is five minutes by default, and a subscription drops to five minutes once you are drawing on usage credits past your plan limit. Step away for longer than that and your next message reprocesses the entire conversation without the cache discount. Run /compact or /clear before long breaks to shrink that hidden cost spike.
Should I use Opus or Sonnet in Claude Code?
Opus 5.5 is the Claude Code default on every paid plan, and at $4 per million input tokens and $20 per million output it costs less than earlier Opus models. Sonnet 5.5 is half that price and handles most coding and bulk work well, so switch to it when a task is high-volume rather than hard. For simple tasks like formatting, brainstorming, and sub-agent work, Haiku 4.5 costs half of Sonnet 5.5 and a quarter of Opus 5.5 per token.