Codex on a ChatGPT plan has a short rolling window that refills within hours and a weekly limit on top. Both count model work, not messages, so long sessions, high reasoning effort, big AGENTS.md files and noisy tool output drain them fastest. New sessions per task, lower effort and compressed context make it last.
How Codex usage limits work
A ChatGPT sign-in draws from your plan: a short window costs you an afternoon, the weekly limit can cost days. An API key has no weekly allowance, only per-minute rate limits.
| Sign-in | What limits you | What happens at the limit |
|---|---|---|
| ChatGPT plan | Short window and weekly limit | Wait for reset or upgrade |
| OpenAI API key | Per-minute rate limits and your budget | Requests slow or fail; spend keeps going |
How to check your Codex usage
Run /status in the Codex CLI at the start of a big task. It shows window and weekly use and when each resets. Your ChatGPT settings show usage across CLI, IDE and cloud tasks.
What burns through your limit
Every turn re-sends the system prompt, AGENTS.md, tool definitions and everything still in context. Four things drive most of it.
- 1Context size
Twenty files read early ride along on every later turn until compaction.
- 2Reasoning effort
Reasoning tokens count. High effort on simple edits wastes allowance.
- 3AGENTS.md size
Sent with every request, plus any nested AGENTS.md files.
- 4Tool output
One test run or repo-wide search can add tens of thousands of tokens.
Seven steps to make it last
- 1One task per session
/newwhen you switch tasks. - 2Compact on purpose
Run
/compactat a natural break. See compact vs clear. - 3Match effort to the task
Use
/model: high effort to plan, lower to implement. - 4Keep AGENTS.md short
Commands, conventions and hard rules. Move the rest to docs.
- 5Keep command output quiet
Failing tests only,
taillong logs. See reducing tool output tokens. - 6Trim MCP servers
Turn off the ones a project does not use.
- 7Write a handoff before the limit
Done, broken, next step. Start the next session from the note.
Limit reached: your options
What to do depends on which limit you hit. If one long session ate the week, fix session hygiene before you upgrade.
| You hit | What to do |
|---|---|
| Short window | Wait a few hours. Review changes and write the handoff note. |
| Weekly limit | Upgrade, switch to an API key, or move to another agent. |
| API rate limit | Slow parallel runs, spread jobs out, or request higher limits. |
Compress what Codex re-sends
The rest of the cost is history the model already used but still receives every turn.
condense compresses that history and never rewrites code, patches or errors. Add a provider in ~/.codex/config.toml; ChatGPT logins still use your plan. See the coding agents docs, our quality evaluation and Codex vs Claude Code.
Frequently asked questions
What does "usage limit reached" mean in Codex?
You used your allowance for the window or the week. Wait for the reset Codex shows, switch to an API key, or upgrade your plan.
When does the Codex weekly limit reset?
On its own cycle, separate from the short window. Run /status in the Codex CLI to see both resets.
Is the Codex CLI rate limit different from the usage limit?
Yes. The usage limit is your plan allowance. On an API key there is no weekly allowance, only per-minute rate limits set by your account tier.
How much Codex do I get on ChatGPT Plus?
OpenAI publishes ranges that change, so check its Codex pricing page. In practice it depends on session length, model and effort.
Does reasoning effort affect my Codex usage?
Yes. Higher effort means more reasoning tokens per turn. Save high effort for planning and hard bugs.
Can I check my Codex usage outside the CLI?
Your ChatGPT account settings show it. Through condense, dense info shows tokens before and after compression.