I fear with the removal of this, I may have to unsubscribe my subscription (not that I want to I muchly enjoy codex.) The already expensive premium plus plan was already too expensive for my blood. The new models make way too many mistakes to consider them good for using and that lack of consistancy is not something I’m willing to pay for.. The cost of tokens on the newer models is way to high for hte mistakes it makes and for the price we pay the limitations put in place are making it a medicore service now to purchase now at best. It feels like we’re getting Netflixed into paying for the higher tiered services, and that is something I personally don’t like.
Removing GPT-5.3-Codex from ChatGPT accounts isn’t platform optimization; it’s a pricing rug pull.
5.3-Codex was the most reliable, token-efficient model for codebase integration, while 5.4/5.5 is still struggling with consistency and token bloat. Moving the stable workhorse model exclusively behind the API while leaving Plus users to deal with the volatile iterations of 5.4/5.5 is anti-consumer.
OpenAI shouldn’t treat its paying subscription base as free beta-testers for unoptimized models while locking proven infrastructure behind a higher paywall.
It is removed from ChatGPT accounts, and only available through API. AKA this means that using codex 5.3 directly with your ChatGPT account whether on plus or pro plan won’t work, and it can only be used through 3rd party apps, such as your mentioned “Cursor” or “Copilot” because those use the API directly with tokens you buy from them, not with your ChatGPT account.
Codex está muito burro. O que aconteceu ? Ele era inteligentíssimo para Codar e entendia de regra de negócio, mas de uns dias pra ka tá pior que o Gemini. Kkkkk
exactly, It is still available in Cursor and others. And that is what I was saying I was doing / using. Using codex 5.3 in Cursor for some things, since I have a high plan paid for on that, and then working in Codex with 5.5 for all else (same project) since I have yet to hit any limits with it. For me this has been working fine. And of course its understood that using it in Cursor is not using my tokens from my ChatGPT/Codex account. Thats obvious as Cursor says that. Also why its been a great combo for me to use them this way. Havent hit limits on either platform and Im getting best of both worlds.
5.5 low blows up credits. Did very simple question to test, 20% gone from the 5h. What model is best to use now? Tried 5.4 - mini, but it made so many mistakes.
I understand making room for other models, but they should at least be better.
5.5 has been very token efficient in my experience so far. no issues of it blowing through in limits or anything. Do you have a lot of skills, plugins, mcps, etc on that may not need to be. Or bloated agent.md file/s? I have seen in some other cases that caused the “experience” 5.5 going through tokens / credits “fast”.
I have also experienced 5.4 to be a suitable temp replacement to 5.3 codex while we wait for the new models to arrive which they made space for. Better is always coming. just have to hold tight, trust the process, and get ready for some epicness coming our way.
Actually, version 5.3 was the most efficient one, and it had become part of my daily routine. Like all good things that just work, you only really notice them once they’re gone.
I think the issue is not so much the removal of 5.3 itself, but rather the fact that in the latest 5.5 version there does not seem to be any configuration option to avoid wasting tokens the way it currently does. I tried changing the reasoning level and switching the speed from fast to standard, but the token consumption was still significantly higher than with 5.3.
I was saddened too when my default 5.3-codex model went away as well. I’m transitioning to 5.4 as my daily for the time being but will also be very interested to see if a codex specific model is every post-trained.
For those that are on API tiers, it’s worth mentioning that Microsoft and other platforms are still offering models in their fleet. Foundry Models sold by Azure - Microsoft Foundry | Microsoft Learn
And if some variant of gpt-5.3-codex ever makes its way into the OSS, I’ll be the first one on board.
I’m jealous of your large H100 collection 
My H100s bring all the vibecoders to the yard.
5.3-codex was WAY WAY better than the current availability to choose from, return the 5.3-codex or the subscriptions will grow to a significantly lower number. I am not impressed from the current availability and not going to continue with this. The current options are not following instructions and they repetitively do the same tasks to eat the tokens, so we can spend more money, guys this is not sustainable profit approach to get more money, ppl will simply switch to other providers (seems atm they are far far better than 5.4 and 5.5 all versions) . Thanks for the understanding.
5.3 codex was perfect instruct model.
5.4 is garbage.
They just want us all to move 5.5 and spend all the tokens faster.
@OpenAI_Support am I right?