Usage limits reference

Claude Code Usage Limits Explained

Claude Code reports several usage windows at once, and they do not behave the same way. Knowing which one is about to stop you is the difference between planning a long task and losing one halfway through.

Published 7 minute read

Limit Lifeboat showing a Claude Code account with separate session, weekly all-models, and model-scoped usage meters.
Claude Code reports several windows at once. The tightest one is the one that stops you.

What windows does Claude Code report?

Anthropic reports usage as a set of windows rather than a single balance. The ones you will see are:

Window Length Covers
Session ~5 hours (300 minutes) Short-term consumption, rolling
Weekly, all models 7 days (10,080 minutes) Everything you run in the week
Weekly, model-scoped 7 days each One model family, tracked on its own

Exactly which windows appear depends on your plan and on what Anthropic currently reports. Some accounts see a scoped window for a heavier model alongside the all-models one; others do not. Treat the set you actually see in /usage as the real one, not a list from an article.

Why the session window does not reset all at once

This is the part that surprises people. The session window is rolling, not a bucket that empties on the hour. Consumption from roughly five hours ago ages out gradually, which means capacity returns in a trickle rather than all at once.

The practical consequence: a heavy burst at the start of a working session keeps pressing on your remaining capacity for the next five hours. If you exhaust the window at 10am, you do not get everything back at 3pm — you get back whatever you spent between 10am and 10:15am at around 3:15pm, and so on. Planning a long task right after a heavy burst rarely works out.

Model-scoped windows change what you should do

When a scoped weekly window is the constraint, you are not out of capacity — you are out of capacity for that model. The all-models weekly window may still have plenty left.

That distinction matters because it opens an option the other windows do not: switching to a lighter model can let you keep working immediately, with no waiting and no account change. Before assuming you are blocked for the week, check which window is actually at zero.

Reading the numbers accurately

/usage inside Claude Code is the authoritative view, and claude.ai/settings/usage shows the same picture in the browser. Anything else, including this project, is a convenience layer over those sources.

Two things make a local reading misleading if you are not careful about them:

  • Age. A number fetched twenty minutes ago is not a live number. Limit Lifeboat marks a reading stale after 30 minutes and shows its age rather than presenting it as current.
  • Which account it belongs to. An inactive saved account can only show its last successful reading. If that account's login has expired, the figure could be hours old.

Pace is the other missing piece. Forty percent remaining means something different depending on whether the last hour consumed two points or twenty. Local history makes that visible; a single percentage does not. The guide to why usage numbers disagree covers which reading to trust.

Planning work around the windows

  • Check before a long task, not during it. The cost of a task is not knowable in advance, but the room you have is.
  • Read the reset time, not just the percentage. Twenty percent with a reset in twenty minutes is a completely different situation from twenty percent with four hours to go.
  • Identify which window is tightest. If it is a model-scoped one, changing model is the cheapest fix available.
  • Expect the weekly window to bite late in the week. Session limits interrupt an afternoon; weekly limits interrupt a Thursday.

If you have already hit a limit, the options at that point are a short list, and waiting is not always the best one. For the equivalent picture on the other provider, see Codex CLI rate limits explained.

Questions and boundaries

How long is the Claude Code session window?

Roughly five hours, and it rolls rather than resetting on a fixed clock. Usage from about five hours ago frees up gradually as time passes, so heavy consumption early in a session keeps weighing on you until that block ages out.

What is the difference between the weekly and the model-scoped weekly limit?

The weekly all-models window covers everything you do across the week. A model-scoped weekly window covers one model family in particular, so a session can be blocked on the Opus window while the all-models window still has room. Dropping to a cheaper model can keep you working when only the scoped window is exhausted.

Can I reset a Claude Code limit early?

No. The windows are enforced by Anthropic and cannot be reset on request. The supported ways to get more capacity are Anthropic's documented extra-usage options and higher plan tiers.

Does checking usage consume any of my limit?

Running /usage does not submit a coding prompt or a model turn, so it does not meaningfully consume your allowance. The same is true of reading usage through the provider's usage services.

Why does my remaining percentage differ between tools?

Different tools read from different sources and at different times. Anthropic's own display is authoritative. A local reading has an age, and a value from an inactive or expired login is a last-known figure rather than a live one.

Free and open source

Know which window is about to run out.

Monitor usage and switch deliberately without replacing the rest of your setup.

Download for Mac