Codex CLI and ChatGPT keep two usage counters: a 5-hour window and a weekly one. They drain separately, and you can hit either one even if the other is barely touched. Plus got its 5-hour limit back recently after a temporary removal, and Pro has none for now. When you hit a limit and the work cannot wait, you have three options: wait for the reset, buy a higher plan, or switch to pay-per-token billing, which has no windows.
How Codex limits work
Codex does not count requests. It measures how much model work you use over a period. Formally there are two counters:
- 5-hour window. Resets every five hours, counted from the first request in the new window.
- Weekly window. One shared budget for 7 days, spent in parallel with the 5-hour one.
Once the 5-hour window is used up, Codex refuses new requests until it resets, even if the weekly limit is nowhere near its ceiling. The reverse happens too: you never run dry on the 5-hour window, yet get refused late in the week because your load was high every day.
Why the limit on GPT-6 Astra runs out faster
Models draw down the budget at different rates. The flagship for complex agent work eats a window much faster than the quick models built for routine jobs. If a session hits the limit after half an hour, the cause is often not the amount of code but the heaviest model chosen where a lighter one would do.
Plus versus Pro
The 5-hour window for Codex and ChatGPT Work on Plus was removed in mid-July 2026, together with the release of GPT-5.6 Sol, and brought back on August 25, 2026. 9to5mac covered the timeline. On both Pro plans ($100, $200) the 5-hour limit is, per the same 9to5mac report, still off; we found no official OpenAI text that says so. The weekly budget on Pro is 5x or 20x that of Plus.
Pro does have limits. The weekly window drains faster on the heaviest models, and the Pro $200 plan (20x) has been temporarily closed to new sign-ups and upgrades since September 10, 2026. Details are in Where to buy Codex now that Pro 20x is closed.
What banked resets are
When people write "Codex limits were reset" or "credits were given out", they almost always mean banked resets: one-off bonus resets of the 5-hour and weekly windows. OpenAI hands them out now and then as a promotion, and they do not become a permanent plan upgrade. The mechanism is described in OpenAI's Help Center: a banked reset opens an extra usage window on top of the normal schedule, but it does not change your subscription plan and does not turn into credits you can buy.
In practice: if you got a banked reset today and the limit ran out again tomorrow, that is expected behavior, not a bug. The one-off bonus is spent, and the regular windows work as before.
What to do when the limit runs out
Wait for the window to reset
The most obvious option, but it does not always work. If the deadline is in an hour and the 5-hour window has just started, waiting is not an option.
Switch to a lighter model
If the task does not need the most powerful model, a fast model inside the same Codex stretches what is left of the window, and you can finish before the reset.
Move up a plan
Going from Plus to Pro multiplies the weekly budget, but Pro costs from $100 a month and the limit does not go away: the ceiling is just higher.
Pay per token instead of a subscription
The local Codex surfaces, including the CLI, officially support signing in with an API key instead of a ChatGPT subscription. Billing then follows the tokens you actually use through API billing, and there are no 5-hour or weekly windows. What runs out is your balance, not time.
Anyone paying for Codex from Russia often cannot attach a foreign card to OpenAI directly. With codexkey the same API-key setup works with a Russian card or SBP (the Russian Faster Payments System), with no ChatGPT subscription and no windows:
# ~/.codex/config.toml
model_provider = "codexkey"
model_reasoning_effort = "medium"
model = "gpt-5.6-sol"
[model_providers.codexkey]
name = "codexkey"
base_url = "https://api.codexkey.online/v1"
env_key = "CODEX_KEY"
The CODEX_KEY variable holds your codexkey key, and the full step-by-step setup is in the Codex guide. Three models are available on the balance: gpt-5.6-sol for hard tasks, terra as the middle option, and luna for fast, high-volume work. Everything is priced per token, not per time window. Balance top-ups start at $1, and nothing is billed by time. For the key setup without a subscription in detail, see Codex CLI with an API key instead of a subscription.
FAQ
When does the Codex limit reset?
The 5-hour window resets every five hours from the first request in the current window, and the weekly one is counted over a rolling seven-day period. There is no fixed day of the week for the reset: the timing follows your own activity.
What are Pro Codex limits and how do they differ from Plus?
On Pro ($100 and $200 plans) the weekly Codex budget is 5x or 20x that of Plus, and the 5-hour window, according to 9to5mac, does not apply for now. On Plus both windows are active: the 5-hour one came back on August 25, 2026 after a summer removal.
Is a banked reset a permanent gift?
No. It is a one-off extra reset of a window, which OpenAI hands out as a promotion in particular periods, for example when a new model rolls out. Once the bonus is used, the regular windows follow the standard schedule.
Can I raise my Codex limit without changing plans?
There is no official way to raise the limit on your current plan: you need an upgrade or a banked reset. The only way around the windows altogether is paying with an API key, where usage is counted in tokens rather than in time.
Why did the limit run out faster than last time on a similar task?
Most likely a heavier model was used, or the task needed more agent steps: reading files, running commands, retries. Window usage depends on the model and the amount of context, and the number of messages matters less.