Codex meters your ChatGPT plan in two ways: a five-hour window and a weekly cap. Which of them applies depends on your plan — and that changed on August 25, 2026, when OpenAI brought the five-hour window back for Plus.
Local runs, cloud chats, automations, code review, and delegated workers all draw on the same allowance. There is no separate “Codex CLI” budget.
What changed on August 25
OpenAI had lifted the five-hour limit for a few weeks. On August 25, 2026 it came back for Codex and ChatGPT Work on Plus. Your weekly allowance is now handed out in five-hour slices: burn a slice early and you wait for the reset, even with plenty of weekly quota left.
Pro ($100 and $200) is exempt for the coming months and keeps only the weekly cap. Thibault Sottiaux, who leads Codex engineering, explained the reasoning:
The 5h limit allows us to smoothen the load on our compute, keeping the plan generous in terms of weekly usage. Plus users are relatively casual and new, but sometimes accidentally exhaust their weekly allowance, creating confusion and poor experience.
The weekly cap did not go anywhere. Every plan still has one.
The five-hour window
The window is rolling, and it covers local messages and cloud chats. OpenAI publishes estimates per model, as local messages per five-hour window:
| Model | Plus | Pro 5x ($100) | Pro 20x ($200) | Business |
|---|---|---|---|---|
| GPT-5.6 Sol | 10–100 | 50–500 | 200–2,000 | 10–100 |
| GPT-5.6 Terra | 25–200 | 125–1,000 | 500–4,000 | 25–200 |
| GPT-5.6 Luna | 250–2,000 | 1,250–10,000 | 5,000–40,000 | 250–2,000 |
| GPT-5.4 mini | 60–350 | 300–1,750 | 1,200–7,000 | 60–350 |
Read those ranges as ranges. A “message” is not a fixed unit — it is whatever tokens your turn actually costs, which is why the low end is ten times smaller than the high end.
What your account actually reports
Codex reads your limits from the same ChatGPT endpoint the CLI uses, and it returns a primary and a secondary window. That is the mechanism behind the plan split, and you can see it in the raw data.
On a Pro account checked on August 27, 2026, the endpoint returned exactly
one window — 604,800 seconds, a week — with the secondary window set to null.
No five-hour meter exists on that account at all. On Plus, the five-hour
window is the one that came back in August.
Two consequences worth knowing:
- A missing five-hour bar is not a bug. If your client shows one meter, your plan has one meter.
- Some models carry their own meters. Preview models — GPT-5.3-Codex-Spark, for instance — report a separate five-hour and weekly pair of their own. You can be blocked on one model while another still runs.
Why you burn it faster than you expect
The meter counts tokens, not messages or minutes. Four things dominate:
Every turn resends the conversation. The model is stateless. A one-line follow-up in hour three pays for every file, diff, and tool result behind it. Cached input is cheaper — 10 credits per million against 100 for fresh input on Sol — but it still counts.
Model choice multiplies everything. Per million tokens, Sol costs 100 in and 500 out; Terra 50 and 300; Luna 5 and 30. Running the same task on Sol instead of Luna is roughly an order of magnitude more expensive against your window. Higher reasoning effort adds more on top — the CLI itself warns that high effort “can quickly consume Plus plan rate limits.”
Everything parallel bills to the same pool. Cloud tasks, automations, code review, auto review, and delegated workers all draw on your plan. Cloud chats run on Sol and can cost more of your allowance than the equivalent local run.
Context you never see rides along. Large files pulled into context, chatty MCP tool results, and a bloated AGENTS.md are re-sent on every request that follows them.
How to check where it goes
In the Codex CLI:
/status— current session configuration and token usage./usage— account usage, and redeeming an earned usage-limit reset.
Both read your ChatGPT account, so they include what cloud tasks spent while you were not looking.
What happens at the limit
Codex stops accepting new work until the window resets. If you hold credits, you keep going: on Plus and Pro, credits cover usage after the included allowance runs out, and Business and Enterprise workspaces spend workspace credits under whatever spend controls the admin set.
Credits are billed against token usage at published rates — 100 credits per million input tokens on Sol, 10 for cached input, 500 for output, with that promotional Sol rate running at least through November 21, 2026.
Getting more usage
- Credits. Plus and Pro can buy credits and keep working without changing plans. They apply only after the included allowance is gone.
- Upgrade to Pro. Beyond the larger allowance, Pro currently has no five-hour window — the whole week is yours to spend whenever you want it.
- Business or Enterprise with flexible pricing. Workspace credits replace fixed rate limits, with spend controls for the admin.
The other lever: spend fewer tokens on the same work
Every option above buys a bigger bucket. The other lever is filling it slower. The meter counts tokens — and most tokens a coding agent sends are context the model did not need for this request.
That is the lever Halv pulls. Halv is a desktop app that compresses every request on your machine before it reaches the model — same model, same answer, roughly half the usage. It works with Codex on the ChatGPT plan you already pay for (and with Claude Code, Kimi, and GLM), and a live savings meter shows exactly how many tokens each session avoided sending. The five-hour window stays five hours; you just fit about twice as much work into it.
Try it free — no card required.
FAQ
Does Codex have a 5-hour limit? On Plus, yes — it returned on August 25, 2026, and hands your weekly allowance out in five-hour slices. Pro ($100 and $200) is exempt for the coming months.
Why does Pro not have the 5-hour limit? OpenAI’s Codex engineering lead said the window smooths compute load while keeping weekly usage generous, and that Plus users sometimes burn a full week’s allowance by accident. Pro kept the weekly cap alone.
How do I check how much Codex usage I have left?
/status for the current session’s token usage, /usage for account usage and
to redeem an earned limit reset.
What happens when I hit my Codex limit? New work stops until the reset. Credits, if you have them, let you continue.
Do cloud tasks and local Codex runs share the same limit? Yes — local messages, cloud chats, automations, code review, and delegated workers all draw on one plan allowance.
Does switching to a cheaper model help? Yes. Luna costs 5 in / 30 out credits per million tokens against Sol’s 100 / 500. But every model draws on the same window, so a cheaper model stretches the budget rather than resetting it.
Sources
Limits and prices above come from OpenAI’s own documentation and public statements, checked August 2026:
- Codex pricing: rate limits, credits, and models
- Using Codex with your ChatGPT plan
- Codex rate card
- Using credits for flexible usage in ChatGPT
- Managing credits and spend controls in ChatGPT Business
- Beyond rate limits: scaling access to Codex and Sora
- OpenAI restores 5-hour Codex and Work limits for ChatGPT Plus users
Limits change, and OpenAI has moved this one twice in two months. If a number
here disagrees with /usage, trust the meter.