The GPT-5.3-Codex model does NOT WORK in PAID ACCOUNT

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #14306
• #14237
• #14190
• #14326
• #14181

Powered by Codex Action

Can confirm

Probably a follow-up from #14181.

Also #14306

Thanks for reporting the issue. We're looking into it. Yesterday we made a change to block access to gpt-5.3-codex and gpt-5.4 for users on the Free plan to ensure we had sufficient capacity for paid users. This should not have affected users on paid plans.

It's not yet clear whether the problem you are experiencing is related to this or is a separate issue. Given the timing, it seems likely that they are related. I'll post here when I have an update to share.

If you're a paid user and are not able to access gpt-5.3-codex or gpt-5.4, could you please post your user ID below? We need just a few user IDs to help our investigation. To access your user ID, go to https://chatgpt.com/codex, click on the icon in the upper right to display the profile menu, then click on the top menu item. This will copy your user ID the clipboard. It should look like user_<randomstring>.

@etraut-openai user-NiQGjbXlgLnC1EIvcdeCJl4L

@dmitrydikun sorry about that!
Can you do /feedback on a recent conversation you have in the extension?

@aibrahim-oai (you guys are amazing)

@aibrahim-oai done, feedback id 019cde12-fca7-7801-9c0d-ba6a85be4432

@aibrahim-oai And I don’t have an option to select the GPT-5.4 model in my VS Code extension. As far as I understand, it should be included in the Plus plan.

@dmitrydikun, yes, gpt-5.4 should be available to Plus subscribers. Are you running the IDE Extension or the Codex app? In either case, please make sure you're using the latest version. It looks like you are on an older version. Was gpt-5.4 available previously?

@etraut-openai I’m using the latest versions of VS Code and the Codex Extension — both were installed yesterday. Initially, I was using a JetBrains IDE with the Codex plugin, and I first encountered this issue in JetBrains after updating the plugin. I assumed the problem was with JetBrains, so I switched to VS Code — but the issue remained. In VS Code, the Codex Extension did not show GPT-5.4 from the start.
VS Code Extension info:

Identifier openai.chatgpt
Version 26.304.20706
Last Updated 1 day ago

@etraut-openai if it helps, for me, this problem was fixed by clearing the auth cache and reauthenticating on codex (i can use 5.4 here) and opencode (5.3 access is now restored).

@esskayesss Thanks, but it didn’t help. I closed VS Code, completely deleted the folders ~/.codex and ~/.vscode/extensions/openai.chatgpt-26.304.20706-linux-x64, reinstalled the extension, and logged in again — the issue persists (GPT-5.4 is missing, and 5.3-codex is not working). Also, now I’m being prompted to upgrade to Plus, even though my Plus subscription is already active.

@dmitrydikun may be u can try ask 5.3 or 5.2 codex to fix it

<img width="597" height="193" alt="Image" src="https://github.com/user-attachments/assets/cef08884-95a8-41a5-b2b3-980854ba591d" />

@dmitrydikun may be u can try ask 5.3 or 5.2 codex to fix it <img alt="Image" width="597" height="193" src="https://private-user-images.githubusercontent.com/151509142/561901389-cef08884-95a8-41a5-b2b3-980854ba591d.png?jwt=eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.eyJpc3MiOiJnaXRodWIuY29tIiwiYXVkIjoicmF3LmdpdGh1YnVzZXJjb250ZW50LmNvbSIsImtleSI6ImtleTUiLCJleHAiOjE3NzMzNDcwNjMsIm5iZiI6MTc3MzM0Njc2MywicGF0aCI6Ii8xNTE1MDkxNDIvNTYxOTAxMzg5LWNlZjA4ODg0LTk1YTgtNDFhNS1iMmIzLTk4MDg1NGJhNTkxZC5wbmc_WC1BbXotQWxnb3JpdGhtPUFXUzQtSE1BQy1TSEEyNTYmWC1BbXotQ3JlZGVudGlhbD1BS0lBVkNPRFlMU0E1M1BRSzRaQSUyRjIwMjYwMzEyJTJGdXMtZWFzdC0xJTJGczMlMkZhd3M0X3JlcXVlc3QmWC1BbXotRGF0ZT0yMDI2MDMxMlQyMDE5MjNaJlgtQW16LUV4cGlyZXM9MzAwJlgtQW16LVNpZ25hdHVyZT0wMGZhODkzMTEzZWRlZTk5YWUxYTEwZWQ4Mzk3MmRlNTRjZWJkMjA2ZGUyMmI5MDY0NjMyY2I3MmRjMDE4NWQ2JlgtQW16LVNpZ25lZEhlYWRlcnM9aG9zdCJ9.KSHlB4P4NitxPJ9ScCJ6K0gSfDWh05wAZWDESwt0QD0">

Did this work?

@chrisrbdev Not exactly the same prompt, but one with a similar meaning — it didn’t help.

i've the same problem

If you're a paid user and are not able to access gpt-5.3-codex or gpt-5.4, could you please post your user ID below? We need just a few user IDs to help our investigation. To access your user ID, go to https://chatgpt.com/codex, click on the icon in the upper right to display the profile menu, then click on the top menu item. This will copy your user ID the clipboard. It should look like user_<randomstring>.

Hello, sorry to bug on this- any update or eta on resolution? Still cannot use since Monday. Please let me know if you would like another user id with paid account for troubleshooting. Thanks

@RMFogarty, we aren't aware of any broad issue here. Just to confirm, are you on a paid subscription plan? Have you tried logging out and logging back in? If that doesn't work, try deleting your ~/.codex/models_cache.json and then relaunch. It might just be a caching issue somewhere. If that doesn't work, provide your user ID and we can look into it.

@RMFogarty, we aren't aware of any broad issue here. Just to confirm, are you on a paid subscription plan? Have you tried logging out and logging back in? If that doesn't work, try deleting your ~/.codex/models_cache.json and then relaunch. It might just be a caching issue somewhere. If that doesn't work, provide your user ID and we can look into it.

Yes, i'm on a paid plan. I just tried logging out again and deleting the cache file. i still get: {"detail":"The 'gpt-5.3-codex' model is not supported when using Codex with a ChatGPT account."}

User:
user-ptnx8RCKbjWSc8GPTkcGAftk

Feedback id: 019ce946-9e6f-7933-bfef-cca4bf136989

@dmitrydikun, it looks like you are on a free plan.

@RMFogarty, the user ID you provided above is on a paid plan, but the logs you uploaded indicate that you're logged in with a different account (different user ID) which is a free plan.

@dmitrydikun, it looks like you are on a free plan. @RMFogarty, the user ID you provided above is on a paid plan, but the logs you uploaded indicate that you're logged in with a different account (different user ID) which is a free plan.

thank you! issue solved. did not realize i had subscribed with a different email

thread ID 019ced80-53c0-7230-bc28-b9131a97b7ff

@M-Server-BOT, you are associated with three different subscriptions — one team subscription and two plus subscriptions. You're likely logged in with one that is expired.

I'm going to close this issue because I don't think we have a bug here. Every reported incident that we've investigated has been explained by use of a Free subscription or an expired subscription.

Hello, I have the same problem—I don't see the Codex 5.3 model listed.

<img width="603" height="86" alt="Image" src="https://github.com/user-attachments/assets/b676dd66-3007-40a9-b822-764258cef2f1" />

The 5.3-codex model has been retired to make more capacity for the newer models. It's still available through the API, but it will no longer be available through ChatGPT subscriptions.

So what’s the best alternative to Codex 5.3 right now? Will Codex 5.4 be released?

If I have a PLUS plan, can I use it via the API without incurring additional costs? Using the tokens included in the PLUS package?

I also think that GPT-CODEX-5.3 was the best coding model available. I don't trust GPT 5.5 at all. Did it really have to be discontinued? I don't see the point in that, especially since it was released in February.
@etraut-openai why... ? ;(

@TimothyAlexisVass From what I've been able to find out online, GPT 5.5 has the same features as Codex 5.3—but it's still a “general-purpose” model, not one strictly designed for code generation.

We created separate "codex" models in the 5.1 through 5.3 timeframe, but we have since unified our models. There is no longer a distinction between codex and non-codex models. Gpt 5.4 and 5.5 are excellent coding models — much better than 5.3-codex.

And they “use up” half as many tokens.

I used GPT 5.5—in a single prompt that changed four lines of code, it used up all the available tokens in the limit.
It took it 10 minutes :) A great alternative—I now understand the point of this.

When OpenAI doesn’t say anything and makes changes quietly in the middle of the night—it means one thing: it’s all about the money :)

<img width="690" height="73" alt="Image" src="https://github.com/user-attachments/assets/54fc2c20-f8b5-42c6-8a8d-ee8daa948a25" />

<img width="551" height="62" alt="Image" src="https://github.com/user-attachments/assets/952a24f2-4f7a-4594-9496-821c8c1afb41" />

When OpenAI doesn’t say anything...

We announced this on x last week.

We regularly retire old models. This is nothing new. It's not practical for us to host old models indefinitely. Almost all Codex users have already transitioned to 5.4 and 5.5.

@etraut-openai Suggesting that your customers' experiences are invalid is called "gaslighting". I advice you to use a GPT model to help you with your responses so that they are less provocative, less defensive, less dismissive and more service oriented. I mean, you could at least use your own product to help you where you clearly have room for improvement...

1. “There is no longer a distinction between codex and non-codex models.”

What you wrote:

“We created separate ‘codex’ models in the 5.1 through 5.3 timeframe, but we have since unified our models. There is no longer a distinction between codex and non-codex models.”

Why this lands badly: the user is saying, “The model I relied on is gone and the replacement feels worse.” Your reply reframes the issue as a taxonomy problem: “There is no distinction anymore.” That does not address the user’s lived experience.

Better version:

“I understand why this is frustrating. Even though we’ve unified the model lineup internally, I realize users may still experience a meaningful difference between 5.3-codex and newer models in real coding workflows. The naming change doesn’t erase that experience.”

1. Gpt 5.4 and 5.5 are better

What you wrote:

“Gpt 5.4 and 5.5 are excellent coding models — much better than 5.3-codex.”

Why this is the most provocative part: the user says “5.5 is worse for me,” and the response is essentially “No, it is much better.” That sounds like telling users their direct experience is wrong, this is called gaslighting.

Better version:

“Our internal evaluations show 5.4 and 5.5 performing better overall on coding tasks, but that may not match every user’s workflow. If you’re seeing worse results, higher token usage, or slower responses, that’s valid feedback and useful for us to investigate.”

1. X is not sufficient as customer communication...

What you wrote:

“We announced this on x last week.”

Why this lands badly: it treats a post on X as sufficient customer communication. Many paying users reasonably expect in-product notice, changelogs, docs, release notes, or deprecation warnings.

Better version:

“We did post about this on X, but I understand that many users may not have seen it. We should have made the change clearer in-product and in the documentation, especially for paid users relying on 5.3-codex.”

1. Retiring of older models

What you wrote:

“We regularly retire old models. This is nothing new.”

Why this sounds dismissive: it minimizes the complaint. For OpenAI, model retirement may be routine. For the customer, it may break their workflow overnight.

Better version:

“We do retire older models periodically, but I understand that this can be disruptive when people depend on a specific model for their coding workflow. We should communicate those retirements clearly and give users a reliable migration path.”

1. Practicality of hosting older models

What you wrote:

“It’s not practical for us to host old models indefinitely.”

This is reasonable as an internal business/technical reality, but it is incomplete. Customers do not mainly care whether indefinite hosting is practical; they care about notice, continuity, and whether they still receive the value they paid for.

Better version:

“We can’t host every older model indefinitely, but we should do a better job explaining the timeline, replacement options, and expected differences before removing a model from ChatGPT subscriptions.”

1. On the topic of common usage

What you wrote:

“Almost all Codex users have already transitioned to 5.4 and 5.5.”

Why this can feel invalidating: even if statistically true, it does not answer the complaint from the users who have not transitioned successfully. It implies, “Most people are fine, so your issue is marginal.”

Better version:

“Many users have moved to 5.4 and 5.5, but that doesn’t mean the transition is working well for everyone. Feedback from users who preferred 5.3-codex is important, especially if the newer models are slower, more expensive in practice, or less predictable for certain coding tasks.”

Just spent way too much time trying to figure out why gpt5.3-codex suddenly doesn't work anymore, double checking subscription and token expiration. Doesn't help that there are multiple replies with the same generic error message that have been dismissed with "PEBCAK". A meaningful error message would go a long way.

@TimothyAlexisVass This is the key passage I think you need to really accept: "It's not practical for us to host old models indefinitely. ". This is the reality. I can assure you that if you, or I, were running their servers, we would be begging our boss to switch off anything but 5.6, even if it's not out yet. Really. I would also focus on being appreciative on having access to an incredible model, which is largely subsidised by OpenAI. Disclaimer: I don't work for OpenAI.

GPT-5.5 burns more than a Ferrari 812 Superfast with a 6.5-liter naturally aspirated V12 engine, 800 horsepower, and a 0–100 km/h time of 2.9 seconds. Maybe I’ll start using Codex again once the limit is refreshed every 5 minutes instead of every 5 hours or once a week.

@mercmobily

This is the key passage I think you need to really accept: "It's not practical for us to host old models indefinitely. ".

You clearly did not read the message. I did not argue with any of the conclusions or the facts, I merely stated the that @etraut-openai has some serious opportunities for improvement in his communication skills.

This is literally what I wrote on that point:

“It’s not practical for us to host old models indefinitely.” This is reasonable as an internal business/technical reality, but it is incomplete. Customers do not mainly care whether indefinite hosting is practical; they care about notice, continuity, and whether they still receive the value they paid for. Better version: “We can’t host every older model indefinitely, but we should do a better job explaining the timeline, replacement options, and expected differences before removing a model from ChatGPT subscriptions.”

Just switch to gpt-5.4-mini with medium reasoning. It's the closest to *-codex models you can get now. I can see you are disappointed, and I DO feel kinda the same, but... what would you expect from OpenAI as a commercial organization? They are still the best value per $ among other subscription-based products.

Could you not have made the error "this model has been retired. Use a different one" ????? Waste of time

Does anyone know, where can I find an up-to-date documentation about available models and pricing? Because gpt-5.3-codex is still (12:26 CEST) listed in the pricing page as "available" for my "Plus" subscription.

The recommended 5.4 is around 30% more expensive based on documentation. I cannot understand, are these numbers correct?

<img width="1124" height="387" alt="Image" src="https://github.com/user-attachments/assets/a4d41a23-ecf4-4b07-a816-1baa0eb10fd1" />

@SemyonSinchenko who needs documentation when the change is announced on some random twitt-ehm-X account 😉

The recommended 5.4 is around 30% more expensive based on documentation. I cannot understand, are these numbers correct?

Jup, this is basically a price increase sold to us as feature

@etraut-openai Please what's solution. I'm experiencing this for the first time, ever since have been using codex

{"type":"error","status":400,"error":{"type":"invalid_request_error","message":"The 'gpt-5.3-codex' model is not supported when using Codex with a ChatGPT account."}}

This is the key passage I think you need to really accept: "It's not practical for us to host old models indefinitely. ". This is the reality. I can assure you that if you, or I, were running their servers, we would be begging our boss to switch off anything but 5.6, even if it's not out yet. Really. I would also focus on being appreciative on having access to an incredible model, which is largely subsidised by OpenAI. Disclaimer: I don't work for OpenAI.

@mercmobily I can accept "It's not practical for us to host old models indefinitely." -- no problems at all. What I cannot accept is "It's not practical for us to update our own documentation with up-to-date pricing and conditions, informing our paid users about changes in terms and limits, as well notify them about breaking changes in their harness setup".

@etraut-openai Please what's solution. I'm experiencing this for the first time, ever since have been using codex {"type":"error","status":400,"error":{"type":"invalid_request_error","message":"The 'gpt-5.3-codex' model is not supported when using Codex with a ChatGPT account."}}

Set the default model to something other that gpt-5.3-codex: https://developers.openai.com/codex/config-basic#default-model

Gpt 5.4 and 5.5 are excellent coding models — much better than 5.3-codex]

Is this your argument?

We were not using this for the benchmarks, but for token efficiency: gpt-5.3-codex that was more token-efficient, more code-efficient, and the true workhorse, intelligent enough to complete 90% of the tasks that 5.5 can, while thinking for fewer minutes and consuming 3x less usage limit per minute spent thinking than 5.5, but has now been removed WITHOUT ANY PRIOR NOTICE BY EMAIL OR IN THE RELEASE NOTES (https://help.openai.com/en/articles/6825453-chatgpt-release-notes).

In addition, the API website (https://developers.openai.com/codex/pricing?codex-usage-limits=plus&codex-credit-costs=plus-pro-enterprise-edu-legacy) still mentioning 5.3 as being available on ALL PLANS.

This is anti-consumer and misleading, leading customers to believe they will have access to the model when they actually will not.

But why they did it?

5.3 is too cost efficient for the user.

After extensive testing in the new OpenAI Codex limit system where it's driven by reasoning time rather than message count. GPT-5.5 is by far the most expensive model in this regard, consuming roughly 3× more limit per minute than GPT-5.3, meaning users get around 20 minutes of reasoning on Plus with 5.5, compared to 60 minutes with 5.3. That difference alone makes GPT-5.3 a much smarter choice for most day-to-day tasks.
What makes GPT-5.3 even more attractive for the user is its behavior, it tends to skip excessive internal reasoning and jump straight into executing the requested changes, which means it not only costs less per minute but also finishes faster (a double efficiency win). The Business plan compounds this problem significantly, as it already drains 3.2× faster than Plus regardless of model and has a 5h our limit that's only 15% of the weekly usage limit, so pairing it with GPT-5.5 can make the limit feel almost unusable for longer tasks.

> @etraut-openai Please what's solution. I'm experiencing this for the first time, ever since have been using codex > {"type":"error","status":400,"error":{"type":"invalid_request_error","message":"The 'gpt-5.3-codex' model is not supported when using Codex with a ChatGPT account."}} Set the default model to something other that gpt-5.3-codex: https://developers.openai.com/codex/config-basic#default-model

Thanks dude! I'd change the 'config.toml' now.

Would be nice if the 5.5/5.4 context compact issue was fixed before 5.3 codex was sunset... Error running remote compact task: stream disconnected before completion: error sending request for url (https://chatgpt.com/backend-api/codex/responses/compact)

I can't use the application successfully at all now, despite trying various workarounds. I get this error about 90% of the time that 5.4 or 5.5 attempts to compact context.

If you're a paid user and are not able to access gpt-5.3-codex or gpt-5.4, could you please post your user ID below? We need just a few user IDs to help our investigation. To access your user ID, go to https://chatgpt.com/codex, click on the icon in the upper right to display the profile menu, then click on the top menu item. This will copy your user ID the clipboard. It should look like user_<randomstring>.

I am plus Member but i am but i am still not able to see the gpt-5.3-codex model can you help me. My email is this: creator@amratpal.com which i used the plus plan on.

@creator-blip If you scroll further down/unfold this thread you will see that gpt-5.3-codex does not exist anymore, see: https://github.com/openai/codex/issues/14331#issuecomment-4599549173

So it does not matter whether you have a plus subscription or not, you will have to switch to another model.

I’ve been forced to use GPT 5.5 since Codex 5.3 was discontinued—after working with it for over a week, I’m just fed up. Here’s a quick question: even a very short and precise prompt can use up all the tokens in 15 minutes.

Do you have an alternative to OpenAI? Claude?

I created my OpenAI subscription because of Codex 5.3. I don't see any particular good reason for continuing the subscription with Codex 5.3 gone and replaced with much less token-efficient "frontier" models.

I recommend Claude Code—forced by OpenAI to look for an alternative, I signed up for Claude.
I’m thrilled. I did buy the PRO subscription, which has 5x more tokens, but I can’t see an end to my work. For comparison, the same tasks on GPT 5.5 took a maximum of 20 minutes. This weekend, I worked in Claude the entire time, and I didn’t even reach 80% usage in 5 hours. I wholeheartedly recommend it!
OpenAI went for the money—it’s already backfired on them, as they’ve fallen behind in the market when it comes to Claude. A shot in the foot.

<uncosntructive_rant>
Jeez, and we're paying premium for this kind of crap. Anthropic pulls the model, Copilot semi-silently hikes the price, OpenAI silently discontinues cost-efficient models..

Another expensive subscription that I am going to now cancel/downgrade, replacing with Chinese providers (kudos to Z.ai).

Do they even feel the need to compete on the market? Or are the IPOs already set as the final rug-pull exits?
</uncosntructive_rant>

Apologies, I am annoyed having to reconfigure my stack yet again this month just to fix it, rather than to improve it.