How the 5-hour rolling window works
Once you send your first message, a 5-hour window opens. Everything Codex does inside it counts toward your allowance: your prompts, the files it reads, tool and terminal output, cloud tasks, and the model's responses. Hit the cap and you are blocked from starting new turns until the window rolls over. If a limit lands mid-turn, Codex is allowed to finish the turn it is on, within fair-use bounds.
The window is shared across surfaces. Local Codex CLI messages, Codex cloud tasks, and ChatGPT Work usage all draw from one pool, so a burst of cloud tasks shrinks what your terminal sessions have left. Regular ChatGPT chat has its own separate allowance.
The weekly limit
On top of the 5-hour window sits a weekly cap sized for sustained use. It exists to stop round-the-clock consumption: you can hit your 5-hour wall with most of the week's allowance still intact, but a few maxed-out days in a row will find the weekly ceiling. When OpenAI paused the 5-hour limit in July, the weekly cap was the limit that remained, which is why "no 5-hour limit" never meant unlimited.
The weekly window resets on its own schedule, roughly seven days from when it opened. During the July capacity crunch OpenAI also reset weekly usage manually several times; those were one-off goodwill resets, not a feature to plan around.
Did Codex get rid of the 5-hour limit? What changed in July 2026
Briefly, yes; permanently, no. The timeline:
- July 9-11: GPT-5.6 (Sol, Terra, Luna) rolls out and demand roughly doubles in 48 hours.
- July 12: OpenAI temporarily removes the 5-hour limit for Plus, Pro, and Business, and resets weekly usage to 100%. The weekly cap stays on.
- July 12-29: repeated weekly resets and efficiency fixes while capacity catches up.
- July 30: the 5-hour limit returns, as announced by the Codex team.
So if Codex feels more constrained now than it did in mid-July, nothing is broken on your end: the temporary window without a 5-hour cap ended. Some Plus and Pro accounts can still earn banked rate-limit resets through referral promotions; redeeming one resets both the 5-hour and weekly windows.
What each ChatGPT plan gets
OpenAI does not publish fixed message caps, because since April 2026 metering is by tokens: how far a window stretches depends on how heavy your requests are. Community measurements exist, but they are estimates and drift with model efficiency changes. What is stable is the relative sizing:
| Plan | Codex capacity | You'll hit the limit when… |
|---|---|---|
| Free & Go | Minimal access, enough to evaluate it. | Almost immediately for real work. |
| Plus | The entry tier for real Codex work; short focused sessions. | Heavy debugging, large file reads, or agentic multi-step runs. |
| Pro (x5 / x20) | Roughly 5x or 20x the Plus allowance; the x20 tier is sized for full-workday use. | Long multi-project days (x5); rarely (x20). |
| Business & Enterprise | Per-seat allowances comparable to Plus by default; workspaces with flexible pricing scale through purchased credits. | Depends on whether your workspace buys credits. |
No tier is unlimited. Higher tiers buy bigger windows, and Enterprise flexible pricing converts the question into a credit budget rather than a cap. The right tier depends on workflow more than seniority: many short, narrow prompts use far less than long debugging sessions with lots of file reads.
How to check your remaining Codex usage
Three ways, from quickest to most continuous:
- In the CLI: run
/statusin a Codex session. It shows how much of the 5-hour and weekly windows you have used. - In ChatGPT: open the usage panel in Codex settings for the same numbers plus any credit balance.
- Continuously: if Codex routes through Headroom, the app tracks both windows from your live traffic, so the menu bar shows how much runway is left before you hit either limit.
Usage limits are not the context limit
Two different meters get conflated. The context window is how much a single conversation can hold before the model starts compacting or forgetting; it resets with every new session and no plan upgrade changes it. The usage limit is how many tokens you can push through Codex across all sessions in 5 hours or a week. The "context used" meter in the CLI is about the former; the limit messages that stop you working are the latter. Confusingly, shrinking what goes into context helps both.
Why you hit limits sooner than expected
Most usage gets eaten by content you did not write: tool output, file content Codex reads on its own, repeated conversation history, and large search results. A single failing build or a few large file reads can use as much of your budget as a long thoughtful prompt. Under token metering this is not a rounding error; it is usually the majority of the bill.
For a fuller breakdown of what burns usage fastest, read the Codex usage guide. The same window mechanics exist on the Anthropic side; if you run both agents, the Claude Code usage limits guide is the mirror of this page. And if your question is why the bill (rather than the limit) feels high, see Why is Codex so expensive?. Coding in the ChatGPT app rather than the CLI? The ChatGPT coding usage limits page frames these same windows app-first, and Codex is now in ChatGPT explains the merger itself.
Stretch your plan instead of upgrading
When you hit a limit, OpenAI offers two paid outs: buy extra credits or upgrade the plan. The third option is sending fewer tokens per request. Headroom intercepts your Codex traffic locally, compresses repetitive logs and boilerplate, and forwards a leaner request: about 50% fewer tokens on the noisy items it compresses, typically 40-50% across a session, so noise-heavy work can stretch the same windows up to twice as far. Same workflow, no account changes.
For the full set of tools and tactics, see the Codex cost guide, including Ponytail, the skill that attacks the output side by making Codex write less code.