Codex Rate Limits Discussion Thread

Codex
558

So, like everyone else, i try to use my weekly limit as well as possible, with a 10 - 15 - 20% daily usage, so it covers the whole week. I use generally 1% prompts, max 2%. Now, with the 5 hour limit, the last propmt in the rate usually fails and the “You’re out of Work usage for now” message comes. Besites the fact that your work gets interrupted, the 1-2% out of the weekly limit gets wasted anyway. That means, in a 2 - 5 hours cicles daily, 2-4% of the weekly limit gets purely wasted, because when the 5 hours pass and the limit gets removed, guess what, you have to promt it again to finish the work it started.

So, putting that in perspective, those 2-4% multiplied by 7 you get… 14 - 28% of your weekly limit wasted on this. This is quite abusive, because i pay for my full fee and quite a lot of what i’m paying is getting wasted on nothing.

It’s also abusive to put the 5 hour limit on everyone, even on people who use the system fairly. If there are abusers amongst us, find a way to stop THEM, not impose limits on everyone. If your network doesn’t handle all the work is put into, improve it, don’t force it on us. Or, at least, let the last prompt from the 5 hour limit finish and show the result, THEN show the message and wait for the reset. Otherwise, it’s just a cheap ripoff.

All the best.

559

This 5-hour limit is unacceptable — users are losing paid usage on unfinished work

I want to express my strong objection to OpenAI’s decision to reintroduce the 5-hour usage limit for Plus users.

This is not merely an inconvenience. It fundamentally changes the value and practical usability of a subscription that customers are already paying for.

I fully understand the need for usage limits. I have no objection to a weekly compute allowance. If I consume my entire weekly allowance, that is my responsibility.

But that is not what is happening here.

With the new 5-hour restriction, I can be prevented from working while the vast majority of my weekly allowance is still available. I recently hit the 5-hour limit while approximately 84% of my weekly allowance was still unused.

How does that make sense?

If I have paid for access that includes a weekly allowance, why am I prevented from using that remaining allowance when I actually need it?

Even worse is what happens to work that is already running.

A task can consume a substantial amount of the 5-hour allowance and then be terminated because the limit is reached before the task has actually completed.

The usage consumed by that failed or unfinished task is not returned.

That means the customer can effectively pay twice for the same work:

The first attempt consumes the allowance but produces no usable completed result.

The customer then has to wait for the limit to reset and run the task again, consuming additional allowance simply to obtain the result that the first allocation of compute was supposed to produce.

I find that completely unacceptable.

If OpenAI stops a task because of its own artificial 5-hour restriction, then the usage consumed by an unfinished task should either be refunded, or the task should be allowed to continue until it reaches a successful completion or at least a safe stopping point.

Anything else means customers can lose part of their paid usage through no choice of their own.

From the customer’s perspective, this can feel functionally indistinguishable from having paid usage taken away: OpenAI deducts the compute, prevents the task from completing, provides no refund of the consumed allowance, and then requires the customer to spend more allowance later to finish the same job.

And all of this can happen while most of the weekly allowance remains untouched.

That is indefensible as a user experience.

The lack of meaningful advance notice makes this even worse.

Professional users build workflows around the capabilities and usage model of a paid service. OpenAI allowed users to work around a weekly allowance, giving us the flexibility to decide when during the week we needed to use our compute.

Then that flexibility was suddenly taken away.

Some of us do not use Work or Codex evenly throughout every day. There may be days when I barely use it at all and other days when I need to process substantial amounts of material. A weekly allowance accommodated that perfectly.

A rigid 5-hour gate does not.

And telling customers that the solution is effectively to move to a dramatically more expensive subscription is not an acceptable answer to a restriction imposed on an existing paid service.

I am not asking OpenAI for unlimited compute.

I am asking OpenAI to let paying customers use the weekly allowance they already have.

At minimum, OpenAI should:

  • Allow users to continue drawing from their remaining weekly allowance after the 5-hour allowance is exhausted.
  • Make the 5-hour protection optional for users who prefer to manage their own weekly compute budget.
  • Never terminate an active task solely because the 5-hour boundary has been reached.
  • Refund usage consumed by a task if OpenAI terminates that task before completion because of a usage-limit boundary.
  • Provide clear advance notice before materially changing the usage conditions of an existing paid subscription.
  • Clearly show users what a task is costing while it is running so they can make an informed decision about whether to continue.

The current system creates the absurd situation where a customer can have 80%+ of their weekly allowance remaining and still be unable to work.

A weekly allowance that I am prevented from using when I need it is not meaningfully a weekly allowance.

And compute deducted for a task that OpenAI itself prevents from completing should not simply disappear.

OpenAI needs to reconsider this policy.

560

Instead of subscribing to the pro version, I chose the plus version. Clearly, I’m not a heavy user. I probably only perform a few tasks each month. Obviously, restoring the 5-hour limit would be almost impossible to implement. Chatgpt/Codex is a tool that helps us work more efficiently, rather than a social software. I don’t use Chatgpt 24 hours a day. I only open and use it when I need to. However, the 5-hour limit now basically prevents me from completing a task completely. I hope a new tool can help me complete my work more efficiently, rather than a situation where I have to wait for several days to finish a task and constantly wait for the quota to refresh.

Perhaps you need to ensure that the server is always accessible during each time period, to make the load more stable. However, it is obvious that this poor 5-hour quota design prevents this revolutionary technology from helping humans perform their work better. Currently, the demand for AI is becoming increasingly demanding. A 5-hour limit is simply not sufficient to complete a complex task.

You should encourage users to use this tool at different times. For instance, ChatGPT can provide different usage rates at different times. This way, users who are using it during peak hours will know that their usage during this period will quickly deplete their weekly quota. Meanwhile, users during off-peak hours can be given more allowances to balance the load. Instead of having users complete a portion of the work every 5 hours limit, which is not efficient.

561

Eww… Please remove the 5-hour usage limit. It’s extremely frustrating and seriously disrupts long coding sessions. It breaks the development flow, interrupts ongoing processes, and forces unnecessary waiting time before work can continue.

For developers working on large or complex projects, this limit significantly reduces productivity. Please consider removing it, increasing it substantially, or providing a way for users to continue coding without these forced interruptions.

562

Dear OpenAI support team, I’d like to share one very big frustration and see if anyone else is encountering the same.

I’ve been using Codex on a Plus plan heavily for the past few weeks. Been working great, I felt like I’ve been on top of all my work and had spare weekly usage left unused every week.

However, ever since you announced the return of the 5-hour session limits, starting on the day of the announcement itself, when I used 5.6 Sol, my weekly limit has been consumed a lot faster, despite I was doing less coding heavy work. So for the past 2-3 days basically.

I’ve barely done any work in the last 2-3 days, and I am already on 53% left of my weekly session.

To be clear - I am not complaining about the 5-hour session window (as limiting as it is), I believe that I am experiencing either an overall reduced weekly session limit, or the 5.6 Sol started consuming a lot more tokens. I haven’t really used Terra much, and Luna seems to be more or less as it was, as it was efficient anyway.

Can you advise please if there were any changes to the weekly or Sol limits. Thanks

563

I am also frustrated on it

565

The 5 hour usage limit is consumed in just 5 minutes when using the Sol model. How are we supposed to work with this restriction??? work for five minutes, then wait 5 hours to continue for another five minutes?

You really need to find a solution. I’ve been subscribed and waiting for these usage-limit issues to be resolved for 2 months. If I’m still unable to work smoothly, this will be my last month as a subscriber.

Claude’s five-hour limit is much more generous and allows us to make real progress. In comparison, the current limit here is extremely frustrating and impractical.

566

That’s really great mission. I agreed.

567

El regreso del límite de 5 horas me parece un cambio muy negativo.

Con el sistema anterior, donde el límite semanal era la principal restricción, podía administrar mi propia cuota según mis necesidades. Si un día necesitaba trabajar durante muchas horas seguidas y otro día prácticamente no utilizaba Work o Codex, podía hacerlo libremente.

Ahora puedo tener todavía 70 %, 80 % o incluso más de mi límite semanal disponible y, aun así, quedar bloqueado simplemente porque alcancé el límite de 5 horas. Eso hace que el límite semanal pierda gran parte de su utilidad.

Entiendo perfectamente que OpenAI necesite establecer límites para controlar el uso de recursos, pero tener simultáneamente un límite semanal y uno de 5 horas resulta demasiado restrictivo para quienes hacemos sesiones largas de trabajo.

Preferiría claramente que regresara el sistema basado principalmente en el límite semanal, permitiendo que cada usuario decida cuándo utilizar su cuota.

Si eliminar el límite de 5 horas no es posible, al menos deberían aumentarlo considerablemente o permitir utilizar parte de la cuota semanal restante para continuar trabajando.

Espero sinceramente que OpenAI reconsidere este cambio, porque la experiencia anterior era mucho más flexible y útil.

568

Tell me. How do all of you REALLY feel?

569

I can’t help but agree – the only way to work more or less continuously is to use Luna Max 5.6. But there’s an obvious problem here: the model is very weak at the local ~32b level. Sol medium is roughly on a par with Luna max, whilst Sol high and above eat through the limits very quickly – on xhigh, 100 per cent of the 5-hour limit is used up in literally 10 minutes.

It’s also VERY unpleasant to hear the news that Pro/Business Premium subscribers don’t have a 5-hour limit, while Go/Plus/Business users have to suffer.

I hope OpenAI hears us—I really loved ChatGPT for its limits, but I’m basically being given a hint that it’s time to switch to another service unless openAI increase those limits.

570

Feature Request: Please remove or fundamentally change the 5-hour Work/Codex limit for Plus users

Hi OpenAI Team,

I am a paying ChatGPT Plus subscriber and I would like to provide feedback on the recently restored 5-hour usage limit for ChatGPT Work/Codex.

My issue is not that usage limits exist. I completely understand that compute-intensive models need limits, and I have no fundamental problem with having a weekly allowance.

The problem is the combination of a weekly allowance and an additional restrictive 5-hour allowance.

If I still have, for example, 80% of my weekly allowance remaining, I do not understand why I am prevented from using that remaining capacity when I actually need it.

The 5-hour limit makes the weekly allowance less useful

A weekly allowance gives the impression that I have a certain amount of Work/Codex usage available to me for the week.

However, the 5-hour limit means that I cannot freely decide when to consume that allowance.

In practice, the weekly allowance is therefore not really a pool of usage that I can manage myself. It is a pool that I am only allowed to consume at a predefined maximum rate.

For many users, especially people working on projects, this does not reflect how real work happens.

I might barely use Work for several days and then have one afternoon or evening where I want to spend several continuous hours working on a website or another larger project.

That is exactly when Work is most useful.

But the 5-hour restriction can stop that workflow even though I still have most of my weekly allowance available.

This directly conflicts with the purpose of Work

Work is particularly useful for larger, multi-step and context-heavy tasks.

Those tasks naturally require iteration:

  • implement a change
  • inspect the result
  • identify a problem
  • correct it
  • test it
  • make another change
  • continue developing the project

This is not something that can always be divided into small sessions several hours apart.

If I am actively developing a website and Work stops me after only a few substantial tasks, telling me to wait until the short-term allowance resets, the entire purpose of having an agent capable of sustained project work is undermined.

The context, momentum and workflow are interrupted precisely when the feature becomes most valuable.

My actual experience

Today I had barely used Work.

I used it for only a handful of requests, including work on a website. Some of those requests were larger tasks, but there were only approximately four major instructions.

My 5-hour allowance was nevertheless exhausted extremely quickly while I still had the large majority of my weekly allowance remaining.

That means I can have plenty of paid weekly capacity available and still be unable to continue a project that I am currently working on.

This makes it very difficult to plan serious sessions around Work.

Private Plus users should not need Pro simply to work continuously

I am already paying for ChatGPT Plus every month.

I am a private user rather than a company, so upgrading to a dramatically more expensive professional plan simply to remove or significantly reduce this kind of interruption is not a reasonable solution for my use case.

There is a very large difference between someone who uses ChatGPT professionally at company scale and a private power user who occasionally wants to spend an entire day working intensively on a personal project.

Plus should have a meaningful place between Free and very expensive professional subscriptions.

If Plus users receive a weekly Work allowance, I believe they should have reasonable freedom to use that allowance when they need it.

Otherwise, a heavy Plus user can effectively be rate-limited in a way that makes the practical difference between a paid account and a much more restricted usage pattern feel significantly smaller during intensive work sessions.

Unused weekly allowance can effectively become wasted allowance

There is another consequence that I think is important.

Because the 5-hour limit prevents intensive usage, users may reach the end of their weekly period with unused weekly capacity.

Not because they did not want to use Work, but because they were prevented from using enough of it at the times when they were actually available to work.

For example:

I may have time today to spend six or eight hours working on my website.

Tomorrow I may not have time at all.

If the system stops me today because of the 5-hour allowance, the fact that my weekly allowance resets later does not help me.

My available time and OpenAI’s quota-reset schedule are two completely different things.

From a customer’s perspective, that means I can pay for a subscription containing a weekly allowance but still be unable to make practical use of that allowance before it expires.

The basic model feels backwards

An analogy would be buying a pack of chewing gum but then being told that, although the entire pack belongs to you, you may only eat one piece every hour.

The problem is not the number of pieces in the pack.

The problem is that someone else is deciding the rate at which you are allowed to use something that has already been allocated to you.

I understand that OpenAI needs to manage infrastructure load and prevent extreme usage spikes.

But from the customer’s perspective, capacity planning should ideally not be solved by making a paid weekly allowance artificially difficult to consume when the user actually needs it.

If the infrastructure cannot support unlimited burst usage, I would strongly prefer a system that gives users more control instead of simply stopping their workflow.

Possible solutions

I would strongly ask OpenAI to consider one of the following:

  1. Remove the 5-hour limit for Plus and use the weekly allowance as the primary limit.

  2. Significantly increase the 5-hour allowance, so that several hours of genuine project work are realistically possible.

  3. Allow users to choose between two modes:

    • protected usage with a 5-hour cap for people who want to avoid accidentally exhausting their weekly allowance;
    • unrestricted weekly-pool usage for users who deliberately want to consume their allowance faster.

  4. Allow users to override the 5-hour limit manually, with a clear warning such as:
    “Continuing will consume your remaining weekly allowance faster.”

That would protect casual users without forcing power users into the same usage pattern.

  1. Introduce rollover or more flexible quota management, so unused capacity caused by short-term limits does not simply disappear.
  2. Provide much clearer usage information, including how much an individual Work task consumed and why.

The third or fourth solution would, in my opinion, be particularly fair.

If the purpose of the 5-hour restriction is to prevent users from accidentally exhausting their weekly allowance, let the user decide.

Warn me. Show me the consequences. Show me my remaining weekly allowance.

But if I knowingly choose to use 50%, 80% or even 100% of my weekly allowance during one intensive working day, that should ultimately be my decision.

Final point

I really like ChatGPT and Work is exactly the kind of feature that makes the service significantly more useful to me.

That is why this limitation is so frustrating.

It does not make me want to use Work more carefully. It makes me question whether I can rely on Work for larger projects at all, because I cannot know whether I will suddenly be forced to stop in the middle of a productive session despite having plenty of weekly usage remaining.

Please reconsider the combination of the 5-hour and weekly limits for ChatGPT Plus users.

A weekly allowance that users can manage themselves would be far more useful than an allowance that exists on paper but cannot necessarily be used when the user actually has the time and need to use it.

Thank you for considering this feedback.

571

I’m extremely disappointed with the return of the 5-hour Codex limit and with how quickly Sol consumes it on Plus.

Today I started with 100% of my 5-hour allowance. One real development task ran for roughly 30–40 minutes and consumed 72% of the entire allowance.

I then started a second task with the remaining 28%. Codex consumed all of it and stopped before completing the task. When my 5-hour window reset, I asked Codex to continue exactly where it had stopped. After roughly another hour of work, the new allowance was already down to about 10%.

In practical terms, two development tasks and roughly 1.5 hours of actual Codex work consumed almost two complete 5-hour allowances.

This is simply not a usable workflow for serious development. I now spend more time waiting for the next window than actually working with Codex.

What makes this particularly frustrating is that this was not my experience with the previous model. I have used Codex extensively for months, and hitting the 5-hour limit used to be very rare. Now a single substantial task can consume most or even all of a window.

I fully understand that usage cannot be unlimited on Plus, and I have no problem with having a weekly quota. But if I still have weekly quota available, I should have reasonable flexibility in deciding when to use it. The current 5-hour restriction makes sustained development extremely difficult.

From a user’s perspective, this also creates very strong pressure to upgrade to a much more expensive subscription simply to maintain a workflow that was previously possible.

I am not asking for unlimited Codex. I am asking for a Plus allocation that is actually usable for sustained development. Please reconsider the current 5-hour limit, increase it substantially, or return to a system where the weekly allowance is the primary constraint.

572

I don’t want more usage. I want control over the usage I already have.

I actually preferred the previous weekly-only limit.

If I decide to burn my entire weekly allowance in three huge development sessions using high reasoning, that’s my choice. If I want to conserve it, I can lower the reasoning level or use lighter models for simpler tasks.

And if I reach 0% of my weekly quota after three days? Fine. Perfect, even. I know I’m done for the week. I can move on to another project, do something else, or simply take a break until the reset.

The 5-hour limit is much more disruptive.

Instead of managing a weekly compute budget, I’m forced into a cycle of:

work → hit the 5-hour limit → wait → continue → hit the limit → wait again.

For long-running projects, this fragments the workflow and constantly keeps the project “almost available”.

It creates absurd behavior too. If my limit resets during the night, I’m incentivized to wake up just to type “continue”, let the agent start working again, and go back to sleep.

A productivity tool should adapt to my working rhythm. My working rhythm shouldn’t have to adapt to its rolling quota window.

I understand why OpenAI may want to protect occasional users from accidentally burning their entire weekly allowance. But that could simply be an optional safeguard.

Give users a choice:

Keep the 5-hour protection enabled, or disable it and use the weekly allowance freely.

I am completely fine with having a finite weekly quota.

I just want to decide whether I consume it gradually over seven days or burn it in three intensive sessions.

Please bring back that freedom.

573

I don’t think they care. Customer service cancel my subscription every time I complain. They don’t even read very well.

574

This is not a complaint about having usage limits. I understand that frontier models are expensive to run, and I am not asking for unlimited usage.

I am asking for something much simpler: let me use the allowance I already have when I actually need it.

Today, I started a single coding task at around 10 AM. It was a large and demanding task, so naturally it consumed usage quickly.

My first 5-hour window was exhausted in less than 30 minutes. The task was not finished.

I waited for the window to reset.

When it finally reset, I continued the same task. Less than 30 minutes later, the 5-hour limit was exhausted again. The task was still not finished.

At that point, I ended up continuing the unfinished work from another Plus account. Again, roughly half an hour later, the limit was exhausted and the task was interrupted.

I returned to the first account and continued again. The same thing happened.

Nearly 12 hours later, I am still trying to complete the same request.

The absurd part is that the actual amount of work required is not 12 hours. If I had simply been allowed to consume my existing allowance without being forced to wait for repeated 5-hour windows, I estimate this task would have taken around 2–3 hours.

So the 5-hour limit did not meaningfully reduce my usage. It transformed a few hours of productive work into an all-day interruption cycle.

Work for 30 minutes. Wait for hours. Resume. Get interrupted. Repeat.

That is not a productive rate limit for long-running coding tasks. It actively destroys momentum and makes large projects disproportionately difficult to work on.

The weekly limit already provides a hard ceiling. If I choose to use a significant portion of my weekly allowance in one productive session, that should be my choice. Some days I barely use Codex at all. On other days, I may need to work intensively on a difficult problem. A weekly allowance makes sense precisely because usage is not evenly distributed across every day and every five-hour period.

The current system creates a strange situation where I can still have weekly usage available, but I am unable to use it when I need it most.

I am not asking OpenAI to give me more usage.

I am asking OpenAI to stop preventing me from using the usage I already have.

At minimum, please consider one of these options:

  • Remove the 5-hour window and rely on the weekly limit.
  • Make the 5-hour limit optional for users who prefer to manage their own weekly allowance.
  • Allow a long-running task to finish instead of cutting it off in the middle.
  • Provide a way to use remaining weekly allowance when the 5-hour window is exhausted.

Right now, the system is especially punishing for the exact kind of work Codex is supposed to be good at: large, complex, long-running software projects.

A rate limit should limit usage, not turn a 3-hour task into a 12-hour day.

576

To everyone reading this, including those who have already shared their opinion in detail or more briefly: if you agree with the many people here about the numerous problems caused by the 5-hour limit in its current form, please say so here, but also consider contacting OpenAI Support and explaining your arguments, the arguments you have read here and agree with, our common request, and the different compromises that have been suggested throughout these comments. Ask them to escalate our requests and arguments internally.

The following are some of the points that come to mind, in no particular order and without claiming this list is exhaustive. I suggest using the ones you agree with most. For example, if you believe the 5-hour limit should realistically be removed altogether, there is little point in emphasizing proposals that would keep it while merely increasing the allowance.

  • One possible solution would be to make the 5-hour cap optional, enabled by default for new Plus users, while also allowing users to control what percentage of their total weekly quota may be consumed within that window. For example, someone could choose to set the limit at 40% or 50% of their total weekly quota.
  • Another possibility would be to remove the 5-hour limit altogether, while displaying an explicit warning showing how much quota has already been consumed, how much remains, and allowing the user to decide whether or not to interrupt the task currently in progress.
  • Regardless of whether an interruption is caused by the 5-hour limit, the weekly limit, or any other usage limit, the agent should always be allowed either to finish its current task or to receive an automatic instruction telling it to stop its work cleanly so that it can be resumed later.
  • The maximum amount that can currently be used within a 5-hour window should be increased substantially. Based on many reports, it has become very difficult to make meaningful progress on something as complex as programming an application, carrying out a demanding Work task, or handling other requests that require a lot of resources over a relatively short period of time.
  • It is also important to take into account that, from the user’s perspective, a specific programming, coding, or Work task cannot necessarily be broken down into as many small work sessions as the current 5-hour restriction effectively requires. This significantly reduces the usefulness of these features.
  • Along the same lines, many users report that depending on the nature of the task, only one or two prompts using the Sol model can be enough to exhaust the entire 5-hour window. This can leave the ongoing work or code in an incomplete, or sometimes even inconsistent, state despite the user still having weekly quota available.
  • It is equally important to take into account that people’s availability varies greatly. Many users cannot work on these projects regularly throughout the week and instead need to concentrate their work into intensive sessions whenever they have the time. The current limit prevents those users from making proper use of Work and Codex.
  • There is also the issue of weekly allowance effectively going to waste for users who were physically unable to use their remaining weekly quota before it reset, not because they did not need or want to use it, but because the 5-hour restriction left them no practical way to do so.
  • This can also strongly encourage Plus users to cancel their ChatGPT subscriptions and look elsewhere.
  • Finally, as some people have pointed out, if the underlying issue is server capacity, then that capacity needs to be increased rather than making users bear the consequences of insufficient capacity. In its current form, the restriction does not merely reduce consumption: for some of the main use cases that make Work and Codex valuable, it can make the service extremely difficult to use effectively.
577

As @sebastien1a already mentioned, please consider contacting OpenAI Support directly so our concerns are heard and taken seriously. You can reach them by email at support@openai.com or through the OpenAI Help Center at https://help.openai.com/ by clicking the chat bubble in the bottom-right corner.

Thank you all for sharing your thoughts and helping bring more attention to this issue.

578

I did. No answer… yet. I hope i will get some answer soon.

579

Yes, and of course, please feel free to keep sharing your own experiences here as well. Don’t hesitate to ask than a human gives a response too, because the AI doesn’t always escalate it.