Agreed. I thinking about moving away.
Note: I am not an OpenAI employee
Submitting such does not guarantee any thing will happen, it is better than nothing.
Thank you very much for your help and for pointing me to the /feedback command. It gave me another technical route beyond the usual support channel.
I submitted feedback from Codex Desktop and received this Feedback ID:
no-active-thread-01a0a6c7-3f50-7420-af59-89ced421ec98
Following the suggestion shown after submitting feedback, I also opened an issue in the official Codex GitHub repository:
Issue #45795 — “Codex Desktop y Work pierden el estado de los proyectos, mezclan archivos y repiten trabajo ya completado”
I included the Codex app version, Windows 10 x64, ChatGPT Plus, OpenAI support case 14819002, a timeline starting on September 8, and screenshots documenting the problems.
The issue is not only fast usage consumption. Codex and Work repeatedly re-process tasks that were already completed, lose the real state of projects, mix files and applications, report tasks as completed when they were not actually applied, and this leads to repeated and unnecessary usage consumption.
I also attached evidence from several different days to show that this is not an isolated incident.
Thanks again for pointing me toward /feedback and GitHub. At least now the issue is also registered through a technical channel.
I got the Luna Reserve after running out of my $200 Pro plan, then dropping another $40 which burned up in about 20 minutes. Luna Reserve went from 100-95% over about an hour, then the same session dropped from about 95 all the way to 0 in the blink of an eye. It’s like I got cleaned out. The session wasn’t even a fully running session, nor do i have any agents running behind it.
Incredibly frustrating, feel l like I’m getting taken for a ride here.
hello everyone, I also have a pro x20 subscription, I am faced with the problem of ultra-fast token debit, exhaustion of limits. I support people with this problem, as I myself work with commercial projects with large amounts of work, where codex helps to cover some of the problems. But in its current state, it’s completely unusable. I will consider alternative options, as unfortunately the business will not wait.
I completely agree with your main point. Reading your post felt almost like reading my own support request. I recently sent OpenAI a detailed, 14-point letter describing practically the same situation: development stops, client commitments remain, and I cannot predict how much longer I will be able to work.
Honestly, what makes this especially discouraging is how many posts in this thread describe almost exactly what I am experiencing.
My main account is on Pro 20x, and I also pay for a second Pro subscription. I purchased the second subscription specifically because I had exhausted the first account’s weekly allowance and needed to continue working. Around September 12, the allowances on both accounts were reset, but they were then consumed unusually quickly again. In practice, I got roughly two days of usable work before facing another wait for the next weekly reset. The screenshots I sent Support on September 15 showed 0% remaining on my main account and 7% on the second.
Before Astra, my normal weekly workload using GPT-5.5 and GPT-5.6 generally fit within the available allowance. I did not have to keep switching models simply to continue working. Now, moving between Astra, Sol, Terra and Luna has become a forced part of the workflow. I almost never use Fast mode with Astra.
At the same time, I genuinely like Astra. I moved from Claude Code to ChatGPT and Codex because I preferred their capabilities and workflow. Astra handles many of my tasks very well, and I want to keep using it. That is precisely why this is so disappointing: the more capable model is here, but I can no longer fit it sustainably into my working week.
This is not a question of why complex tasks require computation. I understand that long context, higher reasoning levels and parallel agents increase consumption. My tasks are not always simple either. But I need to understand how much useful work my paid subscription actually supports, and why that has changed so substantially compared with my previous experience. A remaining-quota percentage, on its own, does not answer that.
The worst part is that I have started hesitating to launch ordinary work tasks. Instead of focusing on the project, I find myself checking the remaining allowance, choosing a lighter model and wondering whether continuing the task will leave me unable to finish it. The unpredictability is slowing development down even before I actually reach the limit.
These are real client projects with contractual deadlines. When development stops for several days, those commitments do not disappear. Delays can mean additional costs, lost income and consequences for client relationships.
I have been contacting Support about this since September 11 and have supplied screenshots and detailed explanations. They have responded, and I appreciate that. But I still do not have a substantive outcome: whether the usage on my accounts has been reviewed, whether it was counted correctly, and what explains the change in practical capacity.
I am not claiming that I have already proven a metering bug. That is exactly why I am asking for an investigation of my specific case, rather than another general explanation that different tasks consume different amounts of resources.
I do not want to leave OpenAI. But I may have to—not because I dislike the product, but because I can no longer reliably depend on it for my work. This is not a threat, and I am not asking for unlimited usage. I need understandable usage accounting, predictable access and a sustainable way to bring client projects through to completion.
Dear OpenAI Support Team,
I am writing to formally contest a recent and drastic depletion of my account credits, which does not align with my actual activity.
Over the past few days, my available credits have been completely swallowed up despite the fact that I executing roughly 10 times fewer tasks/requests.
Given my consistent use setup and significantly reduced activity, this sudden surge in consumption seems completely unjustified and suggests a potential technical error, anomaly update, or underlying issue on the platform’s side.
Could you please urgently review this behaivor and investigate this discrepancy?
I have temporarily paused nd . I look forward to your prompt response and a swift resolution to this matter.
Best regards,
Gonzalo Crozier
The current Codex usage limits and credit pricing need serious reconsideration.
My usage allowance appears to be draining unusually fast, even during what I would consider normal development work. For a tool intended to support sustained coding workflows, hitting usage limits this quickly makes the experience frustrating and difficult to rely on.
The option to purchase additional credits also feels disproportionately expensive. Paying roughly 50% of the cost of the monthly subscription for just 250 additional credits is a significant additional expense.
This is particularly problematic for users in countries such as India, where purchasing power and average incomes are substantially different from markets like the United States. A globally priced subscription or credit system does not necessarily represent the same level of affordability across regions.
I would strongly encourage OpenAI to consider:
- Reviewing how quickly Codex usage limits are consumed.
- Providing clearer visibility into what actions consume credits and at what rate.
- Offering more reasonable additional-credit packages for existing subscribers.
- Introducing regional pricing based on purchasing power.
- Providing higher or more practical Codex allowances for paid subscribers.
Codex is genuinely useful, which is exactly why these limitations are frustrating. Better transparency, more reasonable usage allowances, and regional pricing would make it far more accessible and sustainable for developers outside high-income markets.
Prices are highly dependent on the regions where data centers are located – it’s impossible to meaningfully scale operating costs to global conditions.
Example: You produce something in Europe – you leave the European price to the rest of the world because you incur European production costs. In this situation, your product may be cheap for people in some parts of the world, but expensive for others. ¯\_(ツ)_/¯
However, I fully understand your problem…
Yes, it’s ridiculous, I went through $50 worth of credits in 5 minutes yesterday with 1 agent and no subagents. In the past month, I went through $1800 and over $400 of those were in the past three days, and it just keeps increasing. I used to be able to make a $200 account last a whole day, a few days ago it was drained in a matter of hours, and according to yesterday’s experience, it would now be drained in 20 minutes.
Analyze your data. I found that in my case, 99.985% were fast-labelled purchased credits, even though I explicitly forbid fast models. I’m doing a thorough analysis and will post it soon.
repost
Analyse your data.
I went through $50 worth of credits in 5 minutes yesterday with 1 agent and no subagents. In the past month, I went through $1800 and over $400 of those were in the past three days, and it just keeps increasing. I used to be able to make a $200 account last a whole day, a few days ago it was drained in a matter of hours, and according to yesterday’s experience, it would now be drained in 20 minutes.
Analyze your data. I found that in my case, 99.985% were fast-labelled purchased credits, even though I explicitly forbid fast models. I’m doing a thorough analysis and will post it soon.
So, this payment collection issue is still dragging on, causing trouble for many engineers.
As for the fact that OpenAI has lost its image of nobility…
Why are users who work full time panelized from using codex in one full Saturday? Why do I have to do some work everyday? why can I not use it the way my actual time allows?
because users may have long-horizon work with ChatGPT Codex Computer Use. it’s normal to run several threads parallelly for days or weeks.
This is not about banked resets. It is about automatic hard resets triggered by OpenAI.
A banked reset is useful because I can decide when to redeem it. An automatic hard reset is fundamentally different—especially when it also changes the date of the next regular weekly reset.
Here is a simple example:
- My regular weekly reset is scheduled for Saturday.
- I start the week with 100% and carefully budget that allowance across my workdays.
- By Tuesday, I have used 40%, leaving me with 60% until Saturday.
- On Wednesday, you unexpectedly reset my allowance to 100% and move the next reset to the following Wednesday.
At first glance, this looks as though I have received additional capacity. For professional planning, however, it is not a reliably usable bonus.
Before the intervention, I still had 60% available until Saturday and expected to receive a fresh weekly allowance on Saturday. After the intervention, I have 100%, but that allowance may now have to last until the following Wednesday. The scheduled Saturday reset has disappeared.
That leaves me with two bad options:
- I assume that another unexpected reset will occur, use my allowance more quickly, and risk running out of capacity in the middle of an important project if it does not happen.
- I assume that no further reset will occur and carefully spread the new 100% across the next seven days.
Only the second option is professionally responsible. But it also means that the automatic reset has not created any additional predictable capacity. It is genuinely helpful only if I was already close to zero.
This creates the wrong incentive: users who consume their allowance quickly and speculate on another reset benefit the most. Users who plan responsibly cannot safely make use of the supposed gift.
There is also a second problem. Over the past several months, these automatic resets occurred frequently enough—sometimes once, twice, or even three times within a single week—that many intensive users adapted their real-world workflows to the capacity that was consistently available in practice.
That is exactly what happened in our case. We organized our ongoing development and administrative workflows around the capacity that had been available in practice for months. During the past two weeks, this established practice suddenly changed significantly, without clear communication.
This has not merely affected our workflows in theory. Several ongoing tasks have had to be interrupted, postponed, or continued much more cautiously than planned. Important development work is now sometimes left untouched for days because we need to reserve the remaining allowance for urgent tasks.
We pay for the Pro plan precisely because we use Codex seriously and regularly. Yet we can currently no longer plan with confidence whether a project we start can be completed within the same working week.
We would prefer to continue using OpenAI and Codex. However, if the available capacity remains this unpredictable, we will have to seriously consider moving a larger share of our work to other providers or self-hosted models. This is not a threat. It is a necessary response to the fact that professional workflows require reliable operating conditions.
I am not claiming that promotional resets must continue forever, nor am I claiming any entitlement to unlimited additional capacity. What I am asking for is clear communication and a system that allows professional users to plan their work reliably.
Automatically granted additional capacity should therefore take one of the following forms:
- Bonus capacity that does not change the existing weekly reset date;
- A banked reset that users can redeem when they actually need it;
- Preservation of the unused remainder of the existing allowance;
- Or a clearly published and reliable reset schedule.
During a fundamental transition like this, providing a few additional banked resets to affected Pro users would also be a fair temporary measure. It would allow us to complete work that is already underway while adapting our workflows to the changed capacity.
An automatic reset that replaces the remaining allowance and simultaneously postpones the next regular reset is not a reliable benefit. It turns capacity planning into a gamble.
Please do not confuse this issue with user-controlled banked resets. Banked resets are useful precisely because users can decide when to redeem them. The actual problem is the unpredictable hard reset triggered by OpenAI—and the sudden change to a practice that had been established for months.
I agree with the core point here, but I still find it very hard to believe that OpenAI cannot provide a simple, predictable solution for serious Pro users.
This is not an exotic technical problem.
Professional users need two basic things:
- A transparent usage dashboard showing the real monthly and weekly consumption clearly, including how much capacity remains and how each model affects it.
- A simple paid restart button once the included allowance is exhausted.
That is all.
The workflow should simply be:
Included Pro allowance → allowance exhausted → pay for a restart → continue working.
No uncertainty, no hidden eligibility, no guessing whether another automatic reset will appear, no dependence on extremely expensive credits just to keep working.
For serious Pro users, continuity is not a luxury. It is part of the product.
If I am working on client projects, I need to know whether I can finish the work I start. I cannot organize professional development around unpredictable hard resets, shifting reset dates, unclear consumption rules, or a button that may or may not appear depending on unpublished eligibility criteria.
OpenAI already has the technical mechanisms for resets and now has much better usage analytics.
So the solution seems straightforward:
show the real consumption transparently, show the remaining allowance clearly, and when it reaches zero, allow the Pro user to pay for another full usage cycle.
That would solve the continuity problem almost immediately.
What is difficult to understand is why, after months of the same complaints from serious Pro users, this is still not a standard and predictable part of the product.
Non one knows how your limits work. It’s absolute garbage. There is no transparency on how any of it works except it depletes a lot faster than ever before doing the same tasks. What it used to take a few hours of work before I ever saw a limit message now takes about 45 minutes. How? Absolutely ridiculous. What OpenAI is trying to do is they want people to upgrade to the top tier. Just state that instead of playing this stupid game of appearing to be “user-friendly”.
No this is crazy, Current rates are roughly 250 credits for 20 minutes of processing on 5.6 Sol Extra High thinking. Absolutely insane, exhausted my Pro and the credits I bought afterwards got absolutely evaporated in one prompt. Absolutely crazy, I too would like to know what is going on.
When Astra first became available I was surprised and also a bit flabbergasted at how quickly and well it got stuff done. Usage was high, but it got a lot done also, so I was very happy with it.
That has changed drastically within the past 7 days for me. Right now, even on GPT-6 Astra Light, the 5 hour quota was down 50% by 3 small changes that amounted to changing 20 lines total in code. I am not sure what happened, but it was no positive change that’s for sure.
At this point, the situation with OpenAI Work and Codex has become extremely frustrating.
The main problem is not simply that there are usage limits. The problem is that the usable allowance can be consumed incredibly quickly, sometimes after only a few simple instructions, making Work or Codex effectively unusable for serious professional tasks.
Once the limit is reached, the workflow is interrupted completely. In practice, this means that a task can be left unfinished with no reasonable way to continue, even by switching to a lighter model.
What I find even more concerning is the support experience.
The automated support assistant repeatedly gives generic explanations such as:
-
the task may have been too complex
-
the model used may have been too advanced
-
the context may have been too large
But these answers do not explain cases where the available usage disappears after extremely simple operations or within just a few minutes.
In my case, I have now been waiting for more than a week to speak with a real support representative. The chatbot keeps asking for more screenshots and information, but no actual investigation seems to happen. On several occasions it has even directed me to settings or menu options that do not exist in the current ChatGPT interface.
I have been using ChatGPT since its earliest public releases and have been a paying subscriber for years. I use it mainly for development and professional work, and this is the first time I seriously feel that the usage model is making the product unreliable for that purpose.
The issue also does not seem limited to one subscription tier. Users on both Plus and Pro are reporting very similar experiences.
The question is simple:
If the most advanced models and Work/Codex features cannot realistically be used for sustained professional work without exhausting the quota after a handful of interactions, what exactly is the intended use case?
Users should not be blamed for using the advanced models that OpenAI itself provides and promotes.
What we need is:
-
Transparent information about how usage is calculated.
-
Clear visibility into what actually consumes the quota.
-
Reasonable and predictable limits.
-
A way for support to review abnormal usage on individual accounts.
-
A real escalation path when the automated support system cannot solve the issue.
-
Protection against tasks consuming a disproportionate amount of quota because of a bug or accounting problem.
This is not just about wanting “more usage.” It is about predictability and reliability.
A professional tool cannot consume 50%, 80%, or even 100% of an available usage window after a few basic actions without clearly explaining why.
I genuinely hope OpenAI takes these reports seriously and investigates whether there is currently a technical or accounting issue affecting Work and Codex usage.
Otherwise, many long-term users who depend on these tools for real work will inevitably start evaluating competing platforms.
I am already doing so.