Claude Code is not blocked by one limit. It is blocked by three, and each one has a different fix. The error string in your terminal already tells you which one you hit — most people just do not read it, and then spend hours waiting on a clock that was never the constraint.
Here is the mapping, before anything else:
| What the error says | Which limit | Shared across models | What actually fixes it |
|---|---|---|---|
You've hit your session limit · resets 3:45pm | The short rolling window | Yes | Wait for the printed time. Nothing else |
You've hit your weekly limit · resets Mon 12:00am | The weekly allowance | Yes | Wait for the printed day, or change plan / buy capacity |
You've hit your Opus limit · resets 3:45pm | Per-model cap | No | Switch model with /model — you are working again in seconds |
The rest of this page is the reasoning behind those three rows, and why the popular remedies fail.
The mistake that costs the most
Someone hits the weekly cap on Tuesday, sees an error mentioning a reset time later today, and concludes that waiting a few hours will restore access. It will not. The session window and the weekly window are independent budgets. Clearing the short one does not put a single message back into the long one.
Anthropic's own error reference is explicit that these are separate conditions, and that a model switch only rescues one of the three. If the message does not name a model, the reset timestamp it prints is the only real information you have.
1. The session window — around 5 hours, and rolling
The interval people report most often is 5 hours, which is why "5-hour limit" is the phrase you will see everywhere. Treat the number as a description rather than a mechanism: what defines this window is that it starts from your activity and moves with it. It is not a daily quota that lands at midnight.
A rolling window means what it says. Two people on the same plan hit their ceiling at completely different times of day, because the ceiling is a property of their own usage, not of the clock.
This is why "when does Claude reset" has no single answer. The reset time printed in the error is authoritative — it is issued by Anthropic's servers, not calculated by your client. Trust it over any table you find online, including a table built on an anchor that happens to be common.
2. The weekly allowance — the one that actually surprises people
The weekly cap is where the real confusion lives, because it is a 7-day budget shared across every model. Exhaust it and every model is closed to you at once, no matter which one you were using.
Two consequences worth internalising:
- A session reset will not touch it. The five-hour window can roll over three times in a day while your weekly allowance stays exactly where it was: empty.
- It is not proportional to a single long day. Heavy work in one sitting can consume a surprising share of a seven-day budget, which is why "I only worked two days this week" and "I am out of weekly allowance" are both true statements.
Anthropic does not publish a numeric weekly quota for any plan. Any figure you see in a comparison table is an estimate — usually a careful one, but not a documented number. The practical consequence: you cannot calculate your weekly headroom in advance, so the only reliable move is to watch it as it drains.
3. The per-model cap — the only one /model rescues
The third limit applies to a specific model rather than to your account. If the message names Opus or Sonnet, the ceiling is on that model's requests alone, and moving to a different family restores access immediately.
This is the one case where there is a 5-second fix. It is also the case people most often miss, because a blocked terminal reads like a blocked account. It is not. Run /model, pick something from another family, and carry on.
The general guidance from Anthropic's own documentation is worth having in your head: Sonnet is the intended default for the large majority of coding work, Opus is for the genuinely hard problem, and Haiku is for mechanical edits. Leaving Opus on by default is one of the fastest ways to burn a budget on work that did not need it.
The allowance is shared with more than your terminal
This is the part that makes limits feel arbitrary. Your Claude Code session does not have its own budget. Claude Code, Claude in the browser, the desktop app, and the IDE integration all draw on the same subscription allowance.
So a long research afternoon in the browser genuinely shortens your coding evening. Nothing is broken; you spent the same pool through a different door. Anthropic's documentation states that activity across the tools counts against the same limits, and notes that IDE usage counts too.
If you are trying to explain a weekly cap that vanished faster than your terminal history suggests, this is the first thing to check. It is also the reason two people on the same plan with similar coding habits can have very different weeks.
How to watch all three at once
There is no single dashboard. There are four surfaces, and they do not have equal coverage:
| Surface | Shows | Known limitation |
|---|---|---|
/usage in Claude Code | Session and weekly bars, with reset times | Read from local session history on that machine — could miss usage from other devices |
/status | Remaining allocation | Less detail than /usage |
| Usage ring in the desktop app | Current usage, next to the model picker | Only what the desktop app sees |
| Settings → Usage on claude.ai | Everything, in one place | Not in your terminal |
Anthropic's cost documentation is explicit that /usage figures are approximate and drawn from that machine's local history, which makes the web page the only view that counts all your activity. If you want one number to trust, that is the one.
There is also no early-warning threshold. Claude Code does not warn you at 80%. It works right up until it stops, which is why watching the bar is a habit rather than a nicety.
Four remedies that do not work
- Waiting out a session reset when the weekly cap is gone. The most common and most expensive mistake on this page.
- Upgrading mid-window. Moving from Pro to Max raises the ceiling for future windows. It does not refill a window you have already spent.
- Restarting the terminal, or starting a fresh conversation. Neither touches account-level metering. A new session is still the same session window.
- Switching models for a message that does not name a model.
/modelfixes the per-model cap and nothing else.
FAQ
Is the five-hour window the same as a daily limit? No. It is a rolling window that begins with your activity, not a daily allowance that arrives at a fixed hour. That is why it is described in hours and not tied to a wall-clock time.
Does Claude Code stop gracefully when the limit hits? No. Anthropic's API returns an error when the limit is exceeded, and Claude Code cannot continue without API access. There is no task-completion buffer. If it stops mid-edit, that is the expected behaviour, not a bug — which is the argument for keeping work checkpointed.
Can I see the weekly reset time before I hit the cap?
Yes, in Settings → Usage, and in the /usage view inside Claude Code. The printed value is the reliable one; there is no published universal schedule to compare it against.
Sources
- Models, usage, and limits in Claude Code — Anthropic. Metering, shared allowance, model selection guidance, context commands.
- Claude Code error reference — Anthropic. The exact error strings and what each one means.
- Claude plans and pricing — Anthropic. Plan tiers and what each includes.
If you would rather be told when a vendor lifts caps than refresh a usage page, that is what our alert service does — see how the plans compare or subscribe.
Two related reads: what each usage error actually means covers the same three caps vendor-neutrally, which is worth knowing if you also use Codex or Gemini, and Three kinds of reset separates the part you can compute from the part only the vendor can decide.