Usage limit

A usage limit is the ceiling your plan or API key puts on how much model work you can do in a period. Subscription plans (Claude Pro and Max, ChatGPT Plus and Pro) meter usage in rolling windows, commonly a few hours plus a weekly cap, and pause you when it's spent. API keys meter differently: requests and tokens per minute, plus whatever spend tier your account sits in. Both are set by the model provider, not the harness.

The confusing part is that the limit drains in tokens, and tokens are mostly input tokens. A session with a bloated context window burns the allowance many times faster than a lean one, because every model provider request re-sends the whole history. People who 'barely did anything' and still hit the wall usually had one giant session open all day. And a usage limit is not a context limit: one says stop for now, the other says this conversation can't hold more.

To stretch it: keep sessions short and focused, compact before the window balloons, use a cheaper model for grunt work and the strong one for planning, and keep the prefix cache warm so repeated input is billed as cheap cache tokens. If you're regularly capped, an API key with pay-as-you-go pricing removes the window but makes every token visible on a bill, which is its own kind of discipline.