Codex usage limit reached? What to do when Codex runs out

How Codex usage limits work

As of October 2026, Codex is included with ChatGPT Plus, Pro, Business and Enterprise/Edu plans. Your Codex usage shares the plan's allowance with ChatGPT's Work features. Depending on your plan there are two windows:

• A five-hour window. It sets how much you can use in each stretch. A new window starts with your first message after the previous one ends.
• A weekly window. It caps your total for the week.

You can hit the five-hour limit early even when you have weekly allowance left. OpenAI publishes estimated message ranges per model, but your real number depends on the task.

Why Codex runs out faster than you expect

OpenAI's own guidance is clear: small scripts use a fraction of the allowance, while larger codebases, long-running tasks and long sessions use much more per message. Cloud tasks can use more than local ones, and the strongest models use the allowance fastest. A feature that touches several files, followed by a few rounds of test fixes, is exactly the kind of work that drains a window.

How to check your remaining Codex usage

Open Settings → Usage in ChatGPT or the Codex app. It shows how much of each window you've used and when it resets.

What to do when Codex runs out

1. Wait for the window to reset. Fine for the five-hour window, painful for the weekly one.
2. Buy credits or an instant reset. Eligible Plus and Pro accounts can buy extra credits, or a reset that restores both windows at once. A reset pulls your weekly allowance forward rather than adding to it.
3. Switch to a lighter model. Smaller models stretch the same allowance much further for routine work.
4. Use an API key. Codex can run on an API key, billed at standard API rates.
5. Send the heavy work somewhere else. Let Codex plan and review, and give the implementation to a plugin built for it.

Get more done in every window

• Start a new thread for each unrelated task, so Codex isn't re-reading old context.
• Keep AGENTS.md focused. It's read on every task.
• Point Codex at the exact files and functions involved.
• Put the full requirements in your first message, including how to test them.
• Save the strongest model for design questions and hard bugs.

For more, read 8 ways to make AI coding cheaper.

Hand the heavy lifting to XCR

XCR is a free plugin that works like an AI switchboard for your AI coding tool. It plugs into Claude Code, Codex, Cursor and Grok Build, and takes over the expensive part of the job:

1. Your tool hands the task over. When a request needs more than a small edit (roughly more than 30 lines, or more than one file), it goes to XCR first.
2. XCR does the build. It reads the repository itself, writes the change on a copy of your project, runs your project's own checks (type check, lint, tests) and repairs failures within the task budget.
3. You get a short report. Requirements done, files changed, check results and the exact cost.
4. Your tool reviews and wraps up. It checks the change against your request and finishes anything that needs the internet, such as installing packages.

XCR routes each task to models from OpenAI, Anthropic and Google, so one account covers all of them. Small edits and quick questions stay in your tool, exactly as before.

Pricing: model cost plus 20%, plus $0.02 per task. Credit packs of $5, $10, $25 or $50. No subscription. In our 22-task self-test sample, XCR and GPT-6 Astra each completed 19 tasks, and XCR's estimated cost was 41.8% lower.

For Codex, XCR adds itself as an MCP server in your Codex config and puts one short rule in AGENTS.md. Codex then hands bigger changes to XCR on its own. See how the plugin works.