Codex Rate Limit Resets Can Now Be 'Banked'. A Breakdown of How It Works and What to Watch Out For
Have you ever had Codex suddenly go silent on you in the middle of a late-night refactoring session, saying, "That's it for today"?
I have. And not just once. It always happens right during that focused time before a deadline, when I hit the usage limit exactly. Then, the screen shows, "Resets in X hours." Even if it says three hours, I'm thinking, "But I need to get this working right now..."
The timing of that reset was actually always decided by the system. There hasn't been a mechanism until now where users could choose, "Reset right at this moment."
But in June 2026, OpenAI stepped in to change that.You can now use a reset whenever you like.
It might look like a minor change, but when you think about it, it's packed with interesting points. Today, I want to look at how it works and the points you should know before using it.
⏱️ Resets have changed from something you 'wait for' to something you 'choose'
First, let's clarify the premise. There are two types of usage limits for Codex.
A rolling limit every 5 hours shared by CLI local messages and cloud tasks
A weekly limit for long-term projects
Until now, both were designed to automatically return to zero when the system-determined time arrived. The timing of the recovery was completely out of our hands. Nobody benefits if it resets at 3:00 AM, right?
With the newly introduced Rate Limit Banking, you can now save the reset itself. It's like having a token in your hand that sets the counter to zero the moment you use it, allowing you to trigger it at the exact moment you really need it.
This applies to the Go, Plus, and Pro, and Business plans. Along with the rollout, one free reset is granted initially.
🎁 Increasing through referrals, but a line drawn at 'not for sale'
Furthermore, there is a referral program running for a limited time from June 11 to June 24, 2026. Plus and Pro users can invite up to three friends to Codex, and when the invitee sends their first message, both the inviter and the invitee receive one additional reset.
I thought this part was quite interesting; it's not just a simple friend referral gift, but a design tied to the start of new Codex usage. Since there's a benefit for the referrer, it naturally leads to acquiring new users.
By calculation, Plus and Pro users can save up to four resets: one free plus three from referrals. However, saved resets have an expiration date and will expire 30 days after being granted. You can't hoard them indefinitely, so you need to be careful about that.
By the way, OpenAI has clearly stated that they will not create a mechanism to buy resets for a fee. The only ways to get them are the initial free distribution and the referral program.
🔥 Why now, at this timing?
From here, I'd like to think about it from my perspective.
Actually, for the past few months, there has been a lot of noise surrounding Codex's rate limits. On April 2, the consumption calculation method was changed to the same actual token standard as the API, on April 9 the counting method was reviewed, and on April 28, all paid plans had their limits temporarily reset. This seems to have been a response to protests that continued for several weeks on the developer forum.
Even so, it seems the dissatisfaction didn't subside, with ongoing complaints that the reset timing was unpredictable, the consumption rate was abnormally fast, and the remaining balance was difficult to see.
Around that time, reports emerged that Codex's weekly active users had exceeded 3 million. As the number of users grows, the number of people using coding agents as work tools naturally increases as well. When that happens, the pain of hitting the limit and having your work come to a halt is more than just a minor inconvenience. It's as painful as having your equipment shut down.
I have a habit of viewing AI agents as 'automated factory lines,' and it's easy to understand this recent development if you think about it in those terms. Until now, the factory side completely determined the maintenance timing. This time, it feels like they've handed the workers on the floor a few emergency stop release tickets to use when it really counts.
If they can give users a sense of control for the cost of just a few dollars, I think it's a cheap investment for OpenAI.
Summary: The more understated the feature, the more effective it is.
I think the most interesting part of this update is that some of the control over 'when it recovers' has shifted from the system side to the user side. The AI's performance itself hasn't changed at all, but the user experience should change significantly.
By the way, it hasn't been officially stated yet whether the saved resets also apply to the weekly limit. It seems likely to work for the 5-hour limit, but it's unconfirmed whether it will also roll over to recover the weekly limit. If you're curious, I'd love for you to try it out and see for yourself.
I think those who use AI agents as work tools will truly appreciate features like this that allow you to 'design your own use of time.'
Have you ever had trouble hitting the Codex limit? If so, I'd love it if you could let me know in the comments when that happened.
Production Notes
While writing this article, I realized that there was a similar rate limit controversy on the Anthropic side as well. This episode made me realize that the more AI agents become 'everyday work tools,' the more these minor operational details become quietly important.
Miccell - Once you understand how things work, the world becomes more interesting.