How the 5-hour window works
Your first message opens a 5-hour window. Every token Claude Code reads and writes inside it counts toward your allowance: the prompt, files it opens, terminal output, logs, search results, JSON payloads, tool responses, and its own replies. Conversation history is sent again on every turn, so a long session pays for its early context many times over. When the window is full you are blocked from starting new turns until it rolls over.
The weekly cap
Since August 2025, Anthropic has also capped usage per week. The weekly limit is what stops round-the-clock use: you can hit the 5-hour wall several times with most of the week intact, but a few maxed-out days in a row will reach the weekly ceiling, and that one does not clear in a few hours. If you run out by midweek, the weekly limit post covers the usual causes: sub-agents, bulky tool output, and re-read files.
What each plan gets
Anthropic publishes relative sizes rather than fixed token counts, because how far a window stretches depends on how heavy your requests are.
| Plan | Relative capacity | Suits | You'll hit the limit when… |
|---|---|---|---|
| Pro | Entry allowance. | Short sessions, a few focused hours a day. | Heavy debugging or codebase exploration. |
| Max x5 | About 5× Pro. | Using Claude Code most of the workday. | Long, multi-project days. |
| Max x20 | About 20× Pro. | Long, context-heavy sessions across several projects. | Rarely, unless you run many agents in parallel. |
Workflow matters more than tier. Many short, narrow prompts use far less than long debugging sessions with lots of file reads.
How to check your remaining usage
- In Claude Code: run
/usageto see how much of the 5-hour and weekly windows you have used. - On claude.ai: the usage page under Settings shows the same numbers.
- Continuously: if Claude Code routes through Headroom, the menu bar shows both windows from your live traffic, so you see the wall coming instead of meeting it mid-task.
Why you hit the limit sooner than expected
Most of the window goes to content you did not write: tool output, files Claude Code opens on its own, replayed history, and search results. Debugging from logs, codebase exploration, large JSON payloads, shell-heavy tasks, and repeated file reads are the usual culprits. One failing build can use as much of the window as an hour of careful prompting. Why is Claude Code so expensive? breaks down each of these.
Stretch your plan instead of upgrading
The limit counts tokens, so sending fewer tokens per turn moves it. Headroom intercepts Claude Code's requests locally, compresses repetitive logs and boilerplate, and forwards a smaller request: about 50% fewer tokens on the noisy items it compresses, typically 25-50% across a session. Noise-heavy work can stretch the same plan up to twice as far, with the same workflow and no account changes.
For the other tools that help, see the Claude Code cost guide. If you run Codex as well, the Codex usage limits guide is the OpenAI side of this page.