Hit the Claude limit before the bar reached 100%?
What Anthropic documents about limit warnings and stale bars, what users report about stops near 95% and rounding, and how to check your current figures.
If Claude stopped you while the bar you were watching showed less than 100%, the usual documented explanations are a stale bar (/usage can show figures up to an hour old), a different limit from the one you were watching (the weekly limit or a model limit), or usage on the same account from another device or claude.ai since you last looked. Anthropic does not document a stop below 100%. One user has reported a hard stop at about 95%, and another a one-point difference between two screens; those are reports, not confirmed behaviour. This post separates the documented from the reported and shows how to check your real numbers.
What Anthropic documents
Warnings come before the limit. Claude Code can show a message such as "You've used 85% of your session limit · resets 3:45pm" (Claude Code docs). In the Claude apps, warnings read "Approaching session limit" or "Approaching weekly limit", or "You've used X% of your weekly limit" when a percentage is available (Help Center). A warning is not an error, and you can keep working.
Limits can be adjusted. Anthropic says: "We may adjust usage limits at our discretion to manage capacity and ensure fair access to all users. For example, we may limit your usage with weekly and monthly caps or model and feature usage. We don't guarantee uninterrupted service" (Help Center). So the amount behind "100%" is not a fixed, published quantity.
The bars can be stale. If Claude Code cannot fetch your plan limits (usually because the usage endpoint is rate limited), /usage shows the last bars loaded within the previous 60 minutes with a "Showing last-known usage" note. Press r to retry (Claude Code docs). A bar that was at 80% an hour ago can be well above that now.
A usage message is not always a plan limit. A message such as rate limited until 3:14am in /usage means Claude Code cannot fetch your plan limits until then, "not that your work with Claude is limited" (same page). Separately, "Server is temporarily limiting requests (not your usage limit)" is a short server-side throttle that has nothing to do with your quota, and Claude Code retries it automatically (Claude Code docs).
Two limits count at once. You can be stopped by the weekly limit while the session bar looks healthy, and the other way round. Anthropic notes that a burst "can exhaust the weekly allowance before the session window resets" (same page). If you only watched the session bar, the weekly one may have been the one that hit 100%. See session limit vs weekly limit.
A model-specific limit can stop one model first. On plans with a separate weekly Fable limit, you can be out of Fable while the all-models bar still has room (Fable article). Switching to another model keeps you working; see how the Fable weekly limit works.
Shared use can hide in the totals. The limit bars are account-wide, so chat on claude.ai or your phone counts. The Claude Code /usage breakdown only covers this machine (Claude Code docs). See do chat and Claude Code share limits.
What users report
These come from GitHub issues. They are not Anthropic's statements.
- A user reported a hard stop at about 95% (#64885).
- Another reported a weekly limit banner saying 98% while
/usagesaid 99% at the same moment, and put it down to one figure being rounded down and the other rounded to the nearest point (#105890).
No Anthropic page explains either behaviour, so there is no way to tell from outside whether they are intended or defects. If you hit a stop below 100%, rule out the documented explanations first: a stale bar, a different limit from the one you were watching, or usage elsewhere on the same account.
How to check your current figures
- Retry the usage screen. In Claude Code run
/usageand pressrif you see "Showing last-known usage". - Look at the account, not the machine. Open Settings > Usage (claude.ai/settings/usage) on the web, Claude Desktop or mobile. It shows "Current session" with percent used and time remaining, "Weekly limits" for all models, and Fable if your plan includes it (Help Center). Verified Ultra accounts have no "Current session" bar, because they have no five-hour limit (Help Center). More places to look are in where to see your Claude limits.
- Read the message. The limit message names which limit you hit and when it resets, for example "You've hit your weekly limit · resets Mon 12:00am" (Claude Code docs). Switching models does not help with the session or weekly limit.
- Check the account. If you have several accounts, confirm which one Claude Code is using with
/status, which shows "version, model, account, and connectivity" (Claude Code docs). See how to tell which account Claude Code is using.
A buffer, not a rule
Anthropic documents a small "wrap-up allowance" for Claude Code: if the five-hour limit arrives partway through a response, it "may keep working briefly so it can finish the step in progress" and shows "Usage limit reached · wrapping up". It is "small and capped", counts toward the weekly limit, and is rolling out gradually, so you may or may not have it (Help Center; details in the wrap-up allowance article). It is a different thing from a stop at 95%, but it is a reminder that the edge of a limit is not always a clean line.
If you want a second reading
Headroom is a free Mac menu bar app that shows session and weekly use for each Claude Code account signed in on your Mac. It marks bars red at 90% and checks every five minutes and each time the panel opens, with a refresh button to check now. Use it as a second reading next to Settings > Usage, not as a replacement: it reads the same figures Claude Code shows, so it cannot explain a gap those figures do not show, between checks it can be a few minutes behind, and Anthropic does not publish that source as an official interface. If it cannot read the numbers, it says so instead of guessing. Headroom never sends a prompt, so checking does not use your limits. It is independent and not endorsed by Anthropic.