Two windows, one shared pool
Your first coding message opens a 5-hour rolling window, and everything the agent does inside it counts: prompts, the files it reads, terminal output, and its own responses. On top sits a weekly cap sized for sustained use, so a few maxed-out days in a row can stop you even when the 5-hour window has room. Since April 2026 both windows meter tokens rather than messages, which means a single noisy debugging turn can cost more than a dozen short prompts.
The pool is shared across surfaces: the Codex tab in the app, terminal sessions in the CLI, cloud tasks, and ChatGPT Work all drain the same windows. A burst of cloud tasks shrinks what your local sessions have left.
Chat limits and coding limits are separate
Regular ChatGPT conversation has its own allowance and is not affected by your coding usage. Hitting the Codex cap does not lock you out of chat, and burning a day of chat does not touch your coding windows. If the app says you are limited in the Codex tab, the number that matters is the coding pool, not your chat history.
What each plan gets
No ChatGPT plan has unlimited coding. Plus is the entry tier for real work, Pro buys roughly 5x or 20x the allowance depending on tier, and Business seats carry Plus-comparable allowances unless the workspace buys credits. OpenAI does not publish fixed caps because token metering makes the answer depend on how heavy your requests are. The full plan-by-plan table, the July 2026 limit pause and restoration, and the reset mechanics are in the Codex usage limits guide, which is the deep-dive version of this page.
How to check what you have left
- In the app: open the usage panel in Codex settings.
- In the CLI: run
/statusin a session. - Continuously: route your coding traffic through Headroom and the menu bar shows both windows live, before you hit either.
Make the windows last longer
Most of the pool is consumed by machine-generated context: tool output, file reads, and replayed history, not your prompts. Headroom compresses that noise locally before it reaches the model, about 50% fewer tokens on the noisy items and typically 40-50% across a session, so the same plan stretches up to twice as far. For the full set of tactics, see how to reduce ChatGPT coding costs; for background on the merger that renamed the app, see Codex is now in ChatGPT.