What counts
The official docs are vague on purpose. One question per page: the answer, why, and how each tool differs.
- Does thinking count against the limit?
- Yes. Thinking tokens are billed like output tokens. Extended reasoning produces hidden tokens before the visible answer. Every vendor counts them against the same window as the answer itself, which is why a 'think harder' setting drains a session faster.
- Do subagents count?
- Yes, every subagent call bills the parent account. Subagents are separate model calls with their own context. Three parallel explorers cost roughly three times a single call. They are not free parallelism.
- Does prompt caching reduce usage?
- Cached reads are cheaper but still count. Cache hits are billed at a fraction of a fresh read, but the fraction is not zero. Long sessions that re-send a big context on every turn are still the fastest way to burn a window.
- Does web search count?
- The search itself is usually free; reading results is not. Search results come back as tokens into the context, and the model's summary of them is output. A research-heavy session spends quota on both.
- After a manual reset, does my 5-hour timer change?
- Usually the bar refills but the window clock keeps running. A manual reset refills the quota; it does not start a fresh window. Your next window boundary is still based on your own first message. We record what actually happened for each event in the history.
- How is the weekly limit calculated?
- A 7-day rolling window starting from your first use. It is not calendar Monday. The weekly bar starts when you first use the tool after the previous weekly reset and refills 7 days later. Vendors show the exact reset time in their usage view.
- When does the 5-hour window start?
- At your first message after the previous window closed. Not a fixed clock time. If you start at 09:12, your window ends 14:12 regardless of how much you use inside it. Idle time still counts down.
- Is the limit shared across devices and apps?
- Yes. One account, one set of bars. ChatGPT web, the desktop app, and Codex CLI draw from the same Codex bars. Claude.ai chat and Claude Code share Pro/Max limits. Switching devices does not reset anything.