How to make your Claude Code limits last longer

The short answer

Everything below comes from Anthropic’s Claude Code documentation and help center, checked on the date above. Where the docs don’t say something, we don’t guess. For how the limits themselves work, start with Claude Code usage limits, explained.

Where your allowance actually goes

It helps to know what you’re saving. Anthropic’s cost docs explain the mechanics plainly:

• Every request resends the whole conversation. Each time Claude uses a tool, Claude Code sends another request carrying that history. Prompt caching makes the re-read cheaper, but a one-line question in a session that has been open all day still draws usage for the whole conversation.
• Breaks cost a cache rebuild. On a subscription, the main conversation’s cache lives for one hour. Your first message after a longer break reprocesses everything. Once you’re drawing on usage credits, that lifetime drops to five minutes by default.
• Thinking is output. Extended thinking tokens are billed as output tokens, and the docs say the default budget can be tens of thousands of tokens per request, depending on the model.
• Background work counts. Subagents, agent teammates, scheduled /loop tasks and /compact all send requests of their own.
• It’s one shared pool. On Pro and Max, Anthropic says the limits are shared across Claude and Claude Code. A long chat on claude.ai or in Claude Desktop uses the same allowance as your terminal.

Anthropic’s help center also lists the length and complexity of conversations, the features you use, the model and the effort level as the things that decide how fast you use your allowance. There’s no fixed message count to stretch. You stretch tokens.

Check before you start something big

The cheapest saving is not starting a two-hour refactor with 10% of a session left. Before a big task:

/usage

/usage shows your plan usage bars and when they reset (/cost and /stats open the same screen). On Pro, Max, Team and Enterprise it also shows a breakdown of what counted against your limits: skills, subagents, plugins and individual MCP servers as a share of the total, plus flags for behaviors such as long context or cache misses when one accounts for 10% or more of recent usage. Press d or w to switch between the last 24 hours and the last 7 days. That breakdown is estimated from session history on this machine, so claude.ai and other devices aren’t in it.

Two other places to look:

• Settings → Usage on claude.ai shows progress bars for the five-hour session and your weekly limits, and when the weekly limit resets. It sees usage from every surface.
• The status line. Claude Code can show your five-hour and weekly percentages at the bottom of the terminal. See showing Claude Code limits in the status line, or build one with the status line generator.

If you only want the next reset in your own time zone, paste the limit message into the Claude reset time tool.

Pick the model on purpose

This is usually the biggest lever. Anthropic’s cost guide says Sonnet handles most coding tasks well and costs less than Opus, and recommends reserving Opus for complex architectural decisions or multi-step reasoning. The model docs list the current defaults: on Pro, Max, Team and Enterprise, Claude Code defaults to Opus 5.5. So unless you’ve changed it, you’re on the bigger model.

/model sonnet
/model opus
/model haiku

/model switches the model and saves it as your default for new sessions. With no argument it opens a picker; press s on a row to switch for the current session only. The aliases, as the docs describe them:

Switch at the start, not in the middle

Each model has its own prompt cache. The prompt-caching docs say a switch with /model means the next request reads the entire conversation with no cache hits. Claude Code asks you to confirm while the cache is still warm. The same applies to opusplan: every toggle in and out of plan mode is a model switch. Anthropic’s tip: pick your model and effort at the top of a session, and change as little as possible mid-task. A good rhythm is /clear, then /model, then the new task.

Switching won’t lift a limit you already hit

Session and weekly limits apply across all models. Only a model-specific message (“You’ve hit your Opus limit” or “Sonnet limit”) is fixed by switching to a model outside that family. More in Claude Code usage limit reached.

Leave fast mode off if cost matters

/fast makes Opus up to 2.5 times faster at a higher price per token. On subscription plans it doesn’t touch your plan limits at all: the docs say fast mode is available via usage credits only and draws directly from them, even when you have plan usage left. That’s real money rather than allowance, so turn it on deliberately.

Turn effort down for easy work

Extended thinking is on by default because it helps on complex work. For simple tasks, Anthropic suggests lowering the effort level:

/effort low
/effort medium
/effort status

Levels run from low through medium, high and xhigh to max. The model docs say Opus 5.5, Sonnet 5.5 and Haiku 5.5 default to medium, and most other models default to high. Their guidance in short: low for quick exchanges, small changes and renames; high where edge cases are likely, such as a bug in an existing codebase; xhigh for deeper reasoning at higher token spend. They also warn that max “is prone to overthinking”, so test it before making it a habit. You can also adjust effort with the left and right arrows in the /model picker.

Some details that trip people up:

• You can’t turn thinking off on Opus 5.5, Sonnet 5.5, Haiku 5.5 or the Fable models. They always think; effort controls how much. On other models, /config has a thinking setting.
• MAX_THINKING_TOKENS only applies to models with a fixed thinking budget. Adaptive-reasoning models ignore nonzero budgets, so use effort there.
• On most models, changing effort mid-session rebuilds the cache, like a model switch. The docs say Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 keep the cache on an API key or Claude subscription.

Keep context small

“Token costs scale with context size” is how the cost docs open this topic. These are the moves that keep it small.

/clear between unrelated tasks

Stale context costs tokens on every later message. /clear starts a new conversation with empty context. Run /rename first (or pass a name, /clear auth-refactor) so you can come back with /resume. Anthropic’s docs note that when you want a fresh start, /clear costs nothing.

/compact at natural breaks, with a focus

/compact Focus on the failing tests and the files we changed

/compact summarizes the conversation to free space, and focus instructions tell it what to keep. It isn’t free: it reads the conversation it summarizes. While the cache is warm that costs a fraction of the context size, but after a long break it reprocesses the full history. So compact between tasks, not after lunch, and not mid-task when auto-compaction would do it anyway. You can also put a “Compact instructions” section in your CLAUDE.md.

/rewind instead of arguing with a bad path

If Claude went the wrong way, /rewind (or double-tap Esc) restores the conversation and code to an earlier checkpoint. The caching docs point out that rewinding lands on a prefix that’s already cached, which is cheaper than compacting a long detour.

See what’s taking space

/context

/context draws your current context as a grid and suggests fixes for heavy tools, memory bloat and capacity warnings.

Trim MCP servers and big pastes

• Disable servers you aren’t using with /mcp. The docs add that CLI tools like gh, aws or gcloud are still more context-efficient than MCP servers.
• Don’t paste whole logs. Point Claude at the file and say what to look for, or let a hook filter it. The cost docs show a hook that rewrites test commands to return only failures, cutting “tens of thousands of tokens to hundreds”.
• Use code intelligence plugins for typed languages. One “go to definition” call can replace a grep plus several file reads.

A lean CLAUDE.md

Your project CLAUDE.md loads at the start of every session, so every line is paid for in every conversation. Anthropic’s memory docs say to target under 200 lines per file; longer files cost more context and are followed less reliably. Claude Code warns at startup and in /status when a file is over the recommended length.

• Keep only essentials: build and test commands, conventions, the few things Claude keeps getting wrong.
• Move workflows into skills. A PR-review checklist or migration recipe can live in a skill, which loads only when used.
• Use path-scoped rules in .claude/rules/ for instructions that matter only in part of the codebase. They load when Claude works with matching files.
• Imports don’t save tokens. @path imports help you organize, but the docs say imported files are expanded and loaded at launch too.

Edits to CLAUDE.md mid-session don’t apply until /clear, /compact or a restart. That’s a feature for your cache, and a gotcha if you expected the change right away.

Avoid wasteful loops

The most expensive tokens are the ones spent going the wrong way and then back. Anthropic’s cost guide lists these habits:

• Write specific prompts. “Improve this codebase” triggers broad scanning. “Add input validation to the login function in auth.ts” lets Claude work with minimal file reads.
• Plan first on complex work. Press Shift+Tab to cycle to plan mode, or type /plan fix the auth bug. Claude explores and proposes an approach for your approval before editing. Plan mode itself doesn’t break the cache, because its instructions are appended as messages.
• Course-correct early. Press Esc the moment Claude heads the wrong way, rather than letting it finish.
• Give it a way to check itself. Tests, expected output or a screenshot let Claude catch its own mistakes before you have to ask for a fix.
• Test incrementally. One file, test, continue.

Anthropic’s general best-practices article adds the chat-side version: put the context up front, group related asks into one message and reread your prompt before sending. A clear first message is cheaper than three clarifying rounds.

Subagents, teams and parallel sessions

Spreading work out can save your main context. It doesn’t save your allowance by default.

Subagents: cleaner context, own bill

The docs suggest delegating verbose work like running tests, fetching docs or reading logs to a subagent, so only a summary returns to your conversation. But each subagent sends its own requests, which count toward the same usage limits, and many subagents returning detailed results can fill your context anyway. To spend less:

• Set model: haiku (or sonnet) in a simple subagent’s frontmatter. Without one, a subagent normally uses your main model.
• Or set CLAUDE_CODE_SUBAGENT_MODEL as a default; add CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1 to apply it to every subagent (v2.1.257 or later).
• Ask for short results: “report only the failing tests with their error messages”.

Agent teams: powerful, and roughly proportional to team size

Agent teams are off by default. Anthropic says they use roughly 7 times more tokens than a standard session when teammates run in plan mode, because each teammate is its own Claude instance with its own context. If you use them: Sonnet for teammates, small teams, focused spawn prompts, and shut teammates down when they’re done.

Parallel sessions and loops

Two terminals on two tasks draw from one pool twice as fast. That’s fine when you mean it. What catches people out is the unintended kind: a /loop or scheduled task fires on its interval even while you’re idle, sending your full context each time. The /usage breakdown has a row for the heaviest loops, so check it if your usage climbs while you sleep.

Plan around the five-hour and weekly windows

Pro and Max have a session limit that resets every five hours and a weekly limit, and every request counts toward both. That shapes a sensible rhythm:

• Start big tasks early in a fresh window. Don’t open a migration with 15 minutes and 10% left.
• Batch small tasks at the end of a window. Quick fixes on Sonnet at low effort are a good use of the last slice.
• Watch the weekly bar from midweek. A heavy Monday and Tuesday can use the weekly allowance before session windows ever bite. If /usage shows the weekly bar running ahead of the week, keep tasks smaller until it resets.
• Take breaks at task boundaries. The one-hour cache means coming back after lunch to a giant session costs a full re-read. /clear before you leave, and start the afternoon fresh.
• Remember claude.ai counts too. A long research chat in the browser comes out of the same pool.

Don’t count five hours from your first message to predict the reset: the pages we checked don’t say exactly what starts the clock. The reset time in /usage or in the limit message is the one to trust. See when does Claude Code reset.

When you’re close to the limit

1. Finish small, park big. Wrap up what’s nearly done and write a short note of the next steps, so the next window starts fast.
2. Let Claude Code wait for you. From v2.1.234, in an interactive session signed in with a claude.ai subscription, Claude Code can wait and continue an interrupted task after the reset. /rate-limit-options shows the options.
3. Use usage credits if the work is worth it. /usage-credits opens Settings → Usage on Pro and Max. Credits are billed at standard API rates; set a monthly cap.
4. Switch tools. Codex runs on your ChatGPT plan, a separate allowance from Claude’s. Claude Code reads a repository’s AGENTS.md when there’s no CLAUDE.md, and Codex reads AGENTS.md too, so one shared instruction file makes hopping between them painless. See Claude Code vs Codex limits and how to check Codex usage.
5. Upgrade if it’s every week. Anthropic suggests Pro to Max 5x, or Max 5x to Max 20x, if you consistently hit limits. Compare in Claude Max vs Pro, or check what your usage is worth with the API value calculator.

One trap: if ANTHROPIC_API_KEY is set in your environment, Claude Code uses it instead of your subscription and you pay API rates. /status shows which login method you’re on.

Habits at a glance

Checklist

Common myths

“Switching to Haiku gets me past a session limit.”

Session and weekly limits apply across all models. Only an Opus or Sonnet limit is fixed by switching family.

“/compact is always free.”

It’s a request that reads your conversation. Cheap with a warm cache, expensive after a long break. /clear is the free one.

“Subagents run on someone else’s budget.”

Their requests count toward the same limits as your main conversation.

“Splitting CLAUDE.md into imports makes it lighter.”

Imports load at launch too. Skills and path-scoped rules are what load on demand.

“Max effort is always better.”

Anthropic says max may show diminishing returns and is prone to overthinking.

What changed recently

Sources

Official pages we read on October 17, 2026. If anything here disagrees with them, they win.

• Claude Code docs: Manage costs effectively
• Claude Code docs: How Claude Code uses prompt caching
• Claude Code docs: Model configuration
• Claude Code docs: Commands
• Claude Code docs: How Claude remembers your project
• Claude Code docs: Subagents
• Claude Code docs: Fast mode
• Claude Help Center: Use Claude Code with your Pro or Max plan
• Claude Help Center: How do usage and length limits work?
• Claude Help Center: Usage limit best practices
• Codex docs: Pricing and usage limits
• Codex docs: Custom instructions with AGENTS.md

FAQ