Your links
- ccusage: see your Claude Code usage per day, month and session from your own logs.
- Claude Code Usage Monitor: live terminal dashboard with your burn rate and when you'll hit the limit.
- RTK (Rust Token Killer): shrinks command output before Claude reads it, 60–90% fewer tokens on common commands (their numbers).
- Caveman: a skill that makes Claude answer short and blunt, about 75% fewer output tokens (their number).
- Claude Code docs: Manage costs effectively: Anthropic's own page, where the 10 tricks below come from.
- What is the Ultra plan?: the help center page, with the "Current session" verification check.
- claude.com/pricing: the current plan table (5-hour and weekly limits).
What changed (30-second version)
Max 20x is now Ultra. Same $200 a month. Verified Ultra accounts have no 5-hour limit any more. But every plan still has a weekly limit, and Ultra's weekly limit is what it always was: 10x Pro.
- Pro, $20 a month: 1x every 5 hours, 1x every week.
- Max (was "Max 5x"), $100 a month: 5x Pro every 5 hours, 5x Pro every week.
- Ultra (was "Max 20x"), $200 a month: no 5-hour limit for verified accounts, 10x Pro every week.
The old "20x" only ever counted the 5-hour window. Anthropic's product lead for Claude subscriptions, Mike Overell, said it on X: when weekly limits were added to Max 20x, they were set to 10x Pro, "We could have done better here, and I'm sorry."
Are you verified? Open Claude, go to Settings > Usage. No "Current session" bar means you're verified. It's automatic, based on your account and payment history. Phone verification at checkout doesn't count. Anthropic hasn't said how long it takes.
The 10 tricks
1. Type and press
In Claude Code, shows your plan bars and a breakdown of what's eating your limit: skills, subagents, MCP servers, loops. Each row comes with a fix, like "Trim CLAUDE.md" or "Compact before you step away". When the footer says , press f and Claude Code applies the fixes it can. Needs a recent version: run .
2. between unrelated tasks
Claude Code re-sends your whole conversation with every message, so old context costs you on every turn. Switching tasks? the session, then . Come back later with . itself costs nothing.
3. with a focus, before you take a break
Your first message after a long break misses the cache and re-processes everything. Compact first, and say what to keep:
/compact Keep the code changes and test results
4. Use Sonnet for most work
Sonnet handles most coding well and costs less than Opus. Save Opus for hard architecture and tricky bugs. Switch with . For simple subagents, set .
5. Turn the effort down on easy tasks
Thinking tokens count as output. For simple edits, lower it with (or in ).
6. Keep CLAUDE.md under 200 lines
It loads into every session. Move long workflow instructions (PR reviews, deploys, migrations) into skills, which only load when needed.
7. Turn off MCP servers you're not using
Run to see what's taking space and to switch servers off. Where a CLI exists (, , ), it's lighter than an MCP server.
8. Plan first, stop early
Big task? Press Shift+Tab for plan mode, so Claude agrees the approach before writing code. Heading the wrong way? Hit Esc right away and use (or double-tap Esc) to go back.
9. Be specific
"Improve this codebase" makes Claude read everything. "Add input validation to the login function in auth.ts" makes it read one file.
10. Watch the hidden burners
These use your limit even while you're away:
- and scheduled tasks re-send your full context every time they fire.
- Agent teams use about 7x more tokens than a normal session. Keep them small and shut them down when done.
- Every subagent's requests count too.
The 4 free GitHub tools
- ccusage: reads your local Claude Code logs and shows usage per day, month and session. is a live view. Get it:
- Claude Code Usage Monitor: live terminal dashboard with your burn rate and a prediction of when you'll hit the limit. Get it:
- RTK (Rust Token Killer): sits between Claude Code and your terminal and shrinks command output (git, tests, logs) before Claude reads it. Get it with their install script or Homebrew. Don't , that's a different project.
- Caveman: a skill that makes Claude answer in short, blunt sentences. It only cuts output, and adds about 1–1.5k input tokens per turn. Get it:
The savings numbers come from each project's own README. Your results depend on how you work. Measure with ccusage before and after.
Still ran out?
- Turn on usage credits to keep going past your limit: run in Claude Code, or go to Settings > Usage on claude.ai. Set a monthly spend limit.
- In recent Claude Code versions, Claude can wait for the reset and carry on the task automatically (see ).