Claude 5-hour limit: when it resets, and the weekly cap that can lock you out for days

TL;DR It's not one limit, it's three bars at once. The 5-hour window starts with your first message and renews 5 hours later (a rolling window). On top of it run two weekly limits: one across all models and one for Sonnet only, both resetting at a fixed time assigned to your account. What drains it fastest: the model (Opus costs far more than Sonnet), the effort level, and long sessions. Check your bars in /usage (d/w for 24h / 7 days). Any of the three cuts you off until it resets.

The 5-hour limit doesn't reset at a set time of day: it starts counting from your first message and resets 5 hours later. The weekly cap does have a fixed day and time, set for your account, and it's the one that can lock you out for days. Both run at once, and /usage tells you when each one resets.

I spend many hours a day in Claude Code, and for months I couldn't work out why it sometimes cut me off mid-afternoon. The reason is that there isn't one counter but several clocks at once, each with its own rules. Once you separate them, it stops feeling random.

> /usage

Current session       ████░░░░░░  18% · resets 3:50pm
Current week (all)    █░░░░░░░░░   3% · resets Jun 25, 8pm
Current week (Sonnet) ░░░░░░░░░░   0%

The clocks

1. The 5-hour window (the "session")

It's a rolling window: it starts counting from your first message and renews 5 hours after that start. It's not wall-clock (not "at midnight"), it's "5 hours from when you began." /usage shows how much you've spent and what time the current session resets. It's the cap most people hit on a heavy day.

2. The weekly limits (there are two)

Above the window run two weekly caps, and here's the key difference: they reset at a fixed time assigned to your account, not a rolling one. It even shows the date of your next reset (e.g. Resets Jun 25 at 8pm). Your reset day and time don't change no matter what you do, and you get your full allowance each cycle.

The two bars are Current week (all models) and Current week (Sonnet only): one counts all your usage, the other only your Sonnet usage. That's how it shows on a Max plan; check your own bars for your plan.

Exception: on an Enterprise seat the pool resets on a rolling window, not a fixed time. The above is for Pro and Max.

What drains it fastest

Not every turn costs the same. The heavy hitters:

• The model. Opus burns far more than Sonnet for the same work. Choosing the right model is the number-one lever.
• Long sessions. /usage warns you itself: above ~150k of context, every turn costs more even when cached. /compact between tasks and /clear when you switch topics.
• Long breaks with the session open. Once the cache expires, the first message on your return reprocesses the entire history at full price, and if you compact right then you pay for it twice. Compact before you step away rather than when you get back. And if you reopen a conversation that has already gone cold, on Pro and Max you get a warning first: the resume dialog that shows what an old conversation costs your 5-hour limit.
• Breaking the cache. A cache miss reprocesses the whole prefix at full price: switching models mid-task can cost 10× on that turn. Covered in prompt caching, and if you want to check whether it is happening to you, /cost counts the misses and names the cause of the last one.
• The effort level, attachments, and tools (web search) add up too. To spend less, the 10 habits to save tokens.

The detail that makes the numbers not add up

Your plan is one shared pool: what you spend on claude.ai, the desktop app, and Claude Code all goes to the same place. But the CLI's /usage is computed locally, and it says so itself: "based on local sessions on this machine, does not include other devices or claude.ai." That's why it sometimes "cuts you off" earlier than /usage suggests: the real server-side counter is ahead of the one you see in the terminal.

Reference

Where it fits

• Keep them always in sight, without running /usage: your 5-hour and 7-day window % in the status line.
• If you hit the 5-hour window partway through a response: Usage limit reached · wrapping up, the slice of your weekly limit Claude uses to finish the step instead of leaving it half done.
• If the limit cuts you off mid-task: how to make Claude Code continue on its own when it resets.
• To see your bars: /usage and /stats. And to find out what is eating that cap, with the name of the skill or MCP server and its percentage: the /usage breakdown.
• Why one turn suddenly spikes: prompt caching, and what broke your cache this time.
• How to spend less: 10 habits to save tokens.
• And with those limits clear, which plan actually pays off for how you work.
• Why a styled page (an artifact) costs more tokens: Artifacts in Claude Code.

Official docs: Usage limit best practices · Models, usage, and limits in Claude Code

Requirements

• The plan bars in /usage (session + weekly) require a subscription plan (Pro, Max, Team, or Enterprise). The exact numbers get bumped from time to time, so always check your own. In early October 2026, Anthropic raised 5-hour limits by 20% and gave Pro, Max and Team one full reset to use once, applied from Settings > Usage until October 22, 2026.