Claude Code usage limit: why it runs out and how to make it last
From the work/ 9 min read
Every message resends the whole conversation, so long sessions drain your plan. Four habits from the docs: /clear, /compact, short sessions and the right model.
- Claude Code
- AI
In this article (9)
Your Claude Code usage limit runs out fast because every message resends the whole conversation, not just the last thing you typed. Four habits make it last: /clear when you switch tasks, /compact at a break in the same task, short sessions, and picking the model with /model at the start. All of it comes from the official docs.
Below: what each habit does and costs, how to see how much context you are carrying, and a command cheat sheet. We read the costs page, the prompt caching page, the commands reference, the context window page and the Help Center on November 12, 2026. None of them publishes quotas as numbers, so we did not invent any: the numbers in the first figure are an illustrative example. To decide what to keep when you compact, see AI coding agent context.
Why context spends your Claude Code usage limit
Each time you send a message, Claude Code makes a new API request. The model remembers nothing between requests, so Claude Code sends everything again: the system prompt, the project context, every earlier message and tool result, and your new message. On top of that, each time Claude uses a tool, another request goes out carrying those results.
Prompt caching makes that resend cheaper: what was already processed is read back at the lower cached rate. But the costs page is direct about it: a one-line question in a session that has been open all day still draws usage for the whole conversation. And on the Pro and Max plans, usage counts against a five-hour session limit and a weekly limit, shared with Claude, per the Help Center.
Scroll the figure sideways to see all of it
The figure uses a simple arithmetic model, an illustrative example and not a measurement: 15,000 tokens of base context plus 3,000 per turn, one request per turn. Computed with Node 22, one 20-turn session sends 930,000 tokens and two 10-turn sessions send 630,000, which is 32% less for the same work. The model ignores caching and undercounts tool calls; it shows the shape of the curve, not your real usage.
In a long session, the docs point to these causes of consumption, and each one has a matching habit:
| What raises it | Why | What you do |
|---|---|---|
| Long context | Every request resends the full conversation | /clear between tasks, /compact at breaks |
| Cache expired after a break | The first message afterwards reprocesses all the context | Short sessions |
| Switching models | Each model has its own cache | Pick it at the start |
| Scheduled tasks | They fire while the session is idle and send the whole context | Stop them when done |
| Subagents | Each one makes its own requests | A smaller model for simple tasks |
/clear: a new session between unrelated tasks
Per the commands reference, /clear starts a new conversation with empty context (aliases: /reset and /new). The previous one stays available with /resume. The costs page lists it first among the habits: stale context wastes tokens on every following message, and /clear costs nothing.
It does not start from absolute zero: before you type, the new session loads your CLAUDE.md, auto memory, MCP tool names and skill descriptions, per the context window page. That is one more reason to keep CLAUDE.md short. If you want to come back later, name the session with /rename before you clear.
Use it when you move to unrelated work: you finished the cart fix and now it is the banner's turn. Do not use it in the middle of a task: you lose the thread, and that is what /compact is for.
/compact: summarize without changing task
/compact replaces the history with a summary so you can stay on the same task with less context. It is not free: the costs page warns that compacting a large context is itself a large request, because it reads the conversation it summarizes. The prompt caching page adds nuance. With a warm cache, that request reads the prefix from cache and costs a fraction of what the context size suggests; after a break longer than the cache lifetime, it reprocesses everything uncached, which is why /compact costs the most when you resume an old session.
The practical takeaway, from the same page: pick your model and effort at the start and keep /compact for a natural break between tasks, not for the moment the context is already bursting mid-task. Claude Code also compacts on its own near the window limit; getting there first lets you decide what is kept. You tell it with instructions, and you can pin them in your CLAUDE.md using the docs' own text:
# Compact instructions
When you are using compact, please focus on test output and code changes
# Compact instructionsis the heading the docs suggest for that section of theCLAUDE.mdat your project root.- The next line tells Claude what to prioritize when it compacts: test output and code changes. Adapt it to your project.
For a one-off, pass the instructions to the command, for example /compact keep the changes in components/cart/ and the test commands.
Short sessions: one task, one session
Two reasons from the docs. First, a short session carries less context in every request. Second, the cache expires. The first message after a break longer than its lifetime reprocesses all the context. For the main conversation that lifetime is one hour on a subscription while you are within the usage included in your plan, and five minutes with usage credits, an API key or a cloud provider.
When you pick a session back up with /resume, Claude Code sends the whole conversation again, and if the cache has expired you pay for it with no discount. A conversation you drag along from yesterday is the most expensive one to open. The short routine: finish the task, /rename, /clear, and the next task starts clean. If you left a scheduled task running, stop it when you are done: per the costs page, it fires even while the session is idle and sends the whole context each time.
Scroll the figure sideways to see all of it
Pick the model for the task
The costs page recommends Sonnet for most coding tasks because it costs less than Opus, and reserving Opus for architectural decisions or multi-step reasoning. For simple subagent tasks, the docs suggest model: haiku in their configuration. Effort is a similar knob: lower /effort for simple tasks, because thinking tokens are billed as output.
Timing matters more than the choice. Each model has its own cache, so switching with /model mid-session makes the next request read the entire history with no cache hits. And /model name saves that model as your default for new sessions; to change only the current session, open the picker with /model and press s on the row.
One detail about limits, also from the costs page. When you see "You've hit your session limit" or "weekly limit", that window is shared across all models and switching models does not restore access. With the model-specific messages ("Opus limit" or "Sonnet limit"), moving to a model from another family does let you keep working.
How to see how much context you are carrying
/context draws context usage as a colored grid and suggests optimizations: heavy tools, bloated memory, capacity warnings. /usage shows session cost, your plan's limits and activity stats (aliases: /cost and /stats). On Pro, Max, Ultra, Team and Enterprise plans it adds a breakdown of what counts against your limits, attributed to skills, subagents, plugins and MCP servers, with recommendations for behaviors that account for 10% or more of your recent usage. It is computed from local history on that machine, so it leaves out other devices and claude.ai. The docs also let you show context usage continuously in the status line.
If you just installed a plugin or an MCP server, that is where you see what share of your recent usage is attributed to it. Installing a plugin is now a one-command job, as covered in Claude Code plugin install from a marketplace; it is worth checking /usage afterwards. A command routine for the day:
/context
/usage
/rename cart-discounts
/clear
/model
/compact keep the changes in components/cart/ and the test commands
/contextbefore a long task: you see how much yourCLAUDE.md, memory and tools already take./usagemidway through the day: the plan bars and the breakdown show which habit costs you most./rename cart-discountsnames the current session so you can find it in/resume./clearstarts the next task with empty context./modelwith no argument opens the picker at the start of the task; presssto change only this session./compactwith instructions, at a break in the same task, to keep only what matters.
Scroll the figure sideways to see all of it
Common mistakes
- Chaining tasks without
/clear. The new task drags along, and pays for, the history of the earlier ones. - Compacting when you resume an old session. With no cache, the summary reprocesses everything as fresh input; clear and start another if you do not need that history.
- Switching models mid-task "to try it". The next request reads everything uncached; decide at the start.
- Mistaking the context warning for the usage limit. Per the costs page, a context or auto-compact warning is not a usage limit: the conversation is just nearing the window size.
- Leaving a scheduled task running. It fires while the session is idle and resends the whole context each time.
Frequently asked questions
Does /clear delete my work?
It clears the conversation, not your files. The previous conversation stays available with /resume, and if you named it with /rename you will find it in the picker.
/compact or /clear?
If you are still on the same task and need what was already discussed, /compact. If you are moving to something unrelated, /clear: it costs nothing and does not carry history you will not use.
Does switching models cost more?
The next request reads the whole history without cache, so yes, once. Pick the model at the start and switch only if the task really changes. In the picker, s applies it to this session only.
How do I see how much of my plan is left?
With /usage: it shows your plan's usage bars. The Help Center mentions /status for monitoring what remains. The official pages do not publish quotas in tokens, so any figure for how many messages fit does not come from them.
Checklist
- Your
CLAUDE.mdis short and includes compact instructions. - One task per session;
/renameand/clearwhen you finish. /compactonly at natural breaks, and with instructions.- Model and effort chosen at the start; Sonnet by default, Opus for the complex work.
/contextbefore a long task and/usagemidway through the day.- You do not
/resumean old session if you can start clean. - Scheduled tasks stopped when you are done.
To decide what to keep and when to compact for answer quality, read AI coding agent context. And if you are adding plugins to your workflow, this article covers installing them in one command and what to check first.
Saw this on Instagram? Comment TOKENS on the Reel and I'll send you the guide.
Written by

Published November 12, 2026 · updated November 18, 2026