Verified against my own account on June 22, 2026, and re-checked on August 28, 2026. Banked rate-limit reset is documented by OpenAI, but it is not a permanent feature offered to everyone on identical terms. Availability, limits, and expiry vary by plan, workspace, region, and any promotion in effect. Trust what your own usage screen shows over anything in this article.
Opening Codex, I found something I had not seen before: 2 resets remaining.
Under the usage section of settings, below the 5-hour and weekly windows, there was a count of resets I could trigger myself. Pressing it brought up a confirmation dialog offering to reset my rate limit so I could keep working.
This is a small change with real consequences. Codex rate limits used to be something you waited out. Burn through the 5-hour or weekly window and your options were to wait for the natural reset or, depending on your plan, buy credits. Now the reset can sit in your account until you decide to spend it.
OpenAI calls this banked rate-limit reset. The part that matters most is the one that is easiest to miss: a banked reset expires 30 days after it is granted. This is not a balance you accumulate indefinitely. It is a resource with a deadline.
What actually gets reset
The feature does not increase your usage allowance. It banks the right to reset a rate limit and lets you spend it later.
What gets reset is worth reading carefully. OpenAI’s wording is rate-limit reset, not weekly usage reset. Codex enforces several limits at once: the pricing page says local messages and cloud chats share your plan’s usage allowance, that weekly limits may also apply, and it presents its message estimates per five-hour period. My own screen shows the 5-hour and weekly windows separately.
When I spent a reset on June 22, 2026, both windows came back — the 5-hour window to 99% and the weekly window to 100% on my screen. That was one account on one day, and I would not assume it generalizes. Nothing in the public documentation promises that every user sees the same behavior.
That distinction turns this from a convenience into something that affects how you schedule work. If you want to implement a feature in one evening, or finish an article through to publication, the reset lets you decide when the limit lifts.
Where to find it
In my environment the flow was:
- Open the profile or settings area of the Codex app
- Open the remaining usage view
- Check the 5-hour and weekly windows
- If a reset count is shown, open it
- Confirm the reset
Whether you see this at all depends on your plan, workspace, region, promotion eligibility, app version, and rollout status. If you work only in the CLI, you may not see it the same way. /status shows remaining limits during a CLI session, but the entry point for actually spending a banked reset is the usage summary in the profile menu. Check remaining usage in the CLI, spend resets in the app.
The 30-day expiry is the real catch
A banked reset is usable for 30 days after it is granted, unless different terms apply. So do not hoard them. A reset you were saving for the right moment can quietly disappear.
What makes this uncomfortable is that my screen showed the count of available resets but not the expiry date of each one. There are issues on the openai/codex repository asking for a longer window and for clearer expiry display. Those are user requests, not specifications, but the underlying frustration is reasonable: if something expires, you need to see when.
When to spend one
I do not spend a reset on light work. If the 5-hour window recovers naturally in thirty minutes, waiting is better. Small fixes and short questions can usually be handled by switching to a smaller model, splitting the task, or simply waiting.
Situations where I do spend one:
- The weekly window is exhausted, not just the 5-hour one
- I need to finish something before publishing and cannot stop midway
- Implementation, testing, fixes, and re-verification should run as one stretch
- My own focused time is about to run out
- A banked reset is close to its 30-day expiry
That last one inverts the logic. If a reset is about to expire, spending it on moderately useful work beats letting it lapse. A pre-publication review pass, a rewrite, an end-to-end test sweep, or a docs sync are all better than nothing.
A reset does not make usage unlimited. Consumption still varies with the model, task size, prompt length, AGENTS.md, MCP servers, and image generation. OpenAI’s own guidance for making usage last is to tighten prompts, keep AGENTS.md small, disable unnecessary MCP servers, and use smaller models where they suffice. The reset is insurance for protecting a work session, not permission to be careless.
A checklist before you press it
- Is only the 5-hour window exhausted, or the weekly one too?
- How many hours until the natural reset?
- Is the work ahead a single continuous block?
- Could a smaller model handle it instead?
- Are unnecessary MCP servers or an oversized AGENTS.md inflating usage?
- Is any banked reset approaching 30 days?
- Is this work you genuinely do not want to interrupt?
- Have you decided what counts as done after spending it?
The last item matters most. Without a defined endpoint, a reset drains into unfocused work. For writing I scope it as: pre-publication checks, WordPress deployment, Search Console verification, social handoff, docs sync. For development: spec review, implementation, tests, fixes, re-verification, commit.
One more thing worth watching, reported by users rather than documented by OpenAI: there are open issues on the openai/codex repository describing agents consuming a banked reset without asking, and cases where the remaining count was not visible. If you rely on banked resets, check the count yourself rather than assuming it is untouched.
A reset is better used as a button that finishes work than one that starts it.
FAQ
Is “Codex reset” the same as usage reset?
The UI may say “reset usage” while the documentation says “rate-limit reset.” The accurate reading is closer to the latter: it restores a limit you have hit, rather than erasing your usage history.
Does it only reset the weekly window?
In my test, no. The 5-hour window came back as well. I would not state that as universal. Check your remaining usage screen immediately before pressing: 5-hour window, weekly window, available reset count, and time to natural recovery.
Does this make usage unlimited?
No. The count decreases, banked resets expire after 30 days, and heavy work will approach the limit again afterward.
How I separated fact from noise
For the factual claims I stayed with OpenAI’s own sources: the Codex changelog for when rate-limit reset banking shipped, the pricing page for the 30-day expiry and the treatment of the 5-hour and weekly windows, and the Help Center for how to spend a reset.
Community sources — GitHub issues, Reddit, the OpenAI forum — were useful for understanding what people struggle with, not for establishing behavior. With AI tooling that distinction is worth keeping strict. Screenshots and discussion give you the lead; official documentation and your own measurements give you the facts.
Decide when to spend it, not whether to save it
Banked reset is not a headline feature. For anyone using Codex as a working tool it is a practical one, because rate-limit recovery is no longer something that only happens on the clock’s schedule.
But it expires in 30 days. When you find resets in your account, the question is not how many you have. It is which piece of work each one is going to finish.
References
- How banked Codex resets work | OpenAI Help Center — the 30-day expiry and how to spend a reset
- Codex Referral Promotions | OpenAI Help Center
- Using Codex with your ChatGPT plan | OpenAI Help Center
- Issue #28888: Extend banked Codex reset expiration beyond 30 days — a user request, not a specification
- Codex pricing (OpenAI Developers)
- ChatGPT Codex changelog