Claude Code Limits Explained: Weekly, 5-Hour and Resets

You are forty minutes into a good session, the agent is mid-refactor, and the terminal stops with You've hit your session limit · resets 3:45pm. Nothing warned you, and now you have a half-finished change and hours to wait.

The rules behind that message are not complicated — they are just scattered across a help centre article, a costs page and an errors reference. Here is how the two limits actually work, what drains them fastest, and what your options are when one of them stops you.

The figures below come from a guide whose author checked everything against Anthropic's documentation on 28 October 2026. Limits change often, so treat the docs as authoritative if something has drifted since.

Two limits, one shared pool

On Pro and Max, Claude Code usage is governed by two ceilings at the same time: a rolling five-hour session limit and a weekly limit. The five-hour window is rolling, not a clock-hour reset — hitting it means you are done until the reset time in the message, not done for the day.

The weekly limit is the one that catches people out. According to the errors reference, usage counts against both allowances simultaneously, and a large burst can exhaust the weekly allowance before the session window resets. You can have plenty of room in your five-hour window and still be blocked.

Crucially, it is one pool. The help centre says all activity in Claude and Claude Code counts against the same limits, IDE extensions included. A morning of long browser conversations leaves you less for the afternoon in the terminal.

Model-family messages work differently. You've hit your Opus limit or You've hit your Sonnet limit applies only to that family, and switching to a model outside it keeps you working. The session and weekly limits apply across all models, so switching does nothing there.

What the plans buy, and what nobody publishes

Plan	Price	Usage vs Pro
Pro	$20/month ($17/month billed annually)	Baseline
Max 5x	$100/month	5x
Max 20x	$200/month	20x
API (Console)	Pay per token	No plan window

What you will not find is a token or message count. Anthropic does not publish exact capacity figures for Pro, Max 5x or Max 20x; the help centre describes the structure and says nothing about numbers. Any "you get 45 messages per window" table you have seen is a third-party estimate, and the estimates disagree with each other. The only honest way to size a plan is to run it for a week and watch /usage.

One recent change is worth knowing. A promotion that raised Claude Code weekly limits by 50% ran from 13 May to 13 September 2026. It ended, but part of it stayed: since 14 September, weekly limits are 25% higher than before the promotion. The five-hour limits were never part of it. Older guides citing a 50% boost are out of date.

What drains the allowance fastest

Usage is tokens, not messages — which explains most "I barely did anything and I'm out" stories. Ranked roughly by how hard they bite:

• Long context. The whole conversation is sent with every request, and every tool use sends another request carrying the results. A one-line question in a session that has been open all day still draws usage for the entire conversation.
• Cache misses after a break. The first message after a pause longer than the cache lifetime reprocesses the full context.
• Subagents and agent teams. Subagent requests still draw on your usage, and agent teams use roughly 7x more tokens than a standard session when teammates run in plan mode, because each keeps its own context window.
• Thinking effort. Extended thinking is on by default and thinking tokens bill as output, with a default budget that can reach tens of thousands of tokens per request. On current models it cannot be switched off, so the lever is effort level.
• Background habits and bloat. Scheduled tasks fire on their interval even while a session is idle, each firing sending full context. A long CLAUDE.md loads at every session start and sits in context for every request.

The through-line is that long sessions are the tax, and nearly every other item gets worse as context grows. Worth noting: /compact reads the whole conversation it summarises, so compacting a huge context is itself a large request, while /clear costs nothing.

See where you stand before you get stopped

Run /usage inside Claude Code. On Pro, Max, Team and Enterprise it shows usage bars plus a breakdown of what counts against your limits — attribution by skills, subagents, plugins and individual MCP servers, each as a percentage of the total. Recommendations arrive one per row, each starting with a share of recent usage, in the style of 25% Trim ~/code/api/CLAUDE.md · 49k tokens per request. Press d or w to switch between the last 24 hours and the last 7 days.

Read reset times from the message or from /usage rather than doing the arithmetic yourself; the docs do not spell out when a window starts counting, and the times in the error reference are examples, not a schedule.

The escape hatch has no brakes

There is no documented way to reset a limit early. The documented options are usage credits, switching to a model outside an affected family, moving to a higher plan, or billing through the API.

That last one changes the failure mode entirely, and it is worth understanding before you reach for it. A plan window stops you; a pay-per-token key does not. One engineering studio describes a client who left an automated AI app running unsupervised for a day — jobs triggering AI calls triggering more jobs, with nothing capping concurrency or spend — and took a bill of over $1,000. Nothing broke. Nobody had built the brakes.

The cheapest countermeasure takes five minutes and no code: set a monthly spend limit in your provider console and turn on the hard-limit enforcement. Without that toggle, the limit only sends alerts while requests keep flowing; with it, calls start failing at the number you chose, which is what you want at 3am. That author's framing is worth borrowing — set it to what you can afford to lose in a bad month, not what you expect to spend.

What to do this week

Run /usage once and read the attribution rows before changing anything; the most common wins are a trimmed CLAUDE.md and a forgotten scheduled loop. Build the habit of /clear between unrelated tasks rather than carrying one session all day. If you are consistently blocked on Pro, the documented upgrade path is Max 5x, especially for larger repositories, which carry more context per request.

And if you are considering the API as an overflow valve, set the provider hard limit on the same day you create the key — not after the first surprising invoice.