How Claude subscription limits work
The short version: two windows per account, refilling on their own schedules.
Two windows, not one
A Claude subscription meters Claude Code through two windows at once. The session window opens with your first message and lasts five hours; the weekly window runs on its own seven-day cycle. Each has its own capacity, and hitting either one pauses the work until that window resets.
The two windows are why usage feels confusing from inside a session: you can have plenty of week left and no session left, or the reverse, and the CLI only warns you when you are already close.
When they reset
The session window resets five hours after it opened — not on the clock hour, and not at midnight. The weekly window resets on its own anchor, one moment in the week that is yours alone. Two accounts almost never share an anchor, which is exactly why a second plan helps: when one account is spent, the other is usually mid-refill.
What counts against them
Prompts and the model work they cause count, scaled by model and length. Reading the usage endpoint does not: it is the same source the official clients consult, it carries no prompt, and it runs no model. Checking your limit does not spend your limit.
Watching it without opening anything
Softcap polls that endpoint with your own credentials and puts every account's two windows in the menu bar — with a countdown to the nearest reset, notifications at 80% and 95%, and widgets if you want them on the desk. Live readings, straight from the endpoint, for as many accounts as you have.
The other provider
Running Codex too? Its limits behave differently: how Codex limits work