This is crazy! i had just signed up Gpt Plus only for Codex, now there’s no option besides moving to Claude.
@VeitB Thanks for sharing that. So it is expected.
I wonder if we would have 5.5 codex or later codex models? The reason I’m raising this is because if you have your spec and plan, personally I feel codex models do exactly what your spec says, where as the new 5.4 and 5.5 generic models trying to be smart and add things, take short cuts that I do not want the code to have. Not saying this is bad for everyone, maybe people start with absolutely no spec can benefit from this behaviour, but definitely not me.
Even if it’s not codex models, is it possible to make some models that do things more deterministicly, i.e. follow instruction more closely and get off rail less?
Finally, I don’t use any social media and if that’s the sole place openai communicate with customers, I would really appreciate this to be raised to management and to have official ways to notify customers of this changes e.g. by emails.
Thanks in advance.
I wanted to share my genuine disappointment with the removal of GPT-5.3-Codex from the Codex platform as of June 2nd.
While I understand the reasoning around fleet simplification, this change has a real impact on day-to-day usability — especially for Plus subscribers. GPT-5.5 consumes roughly 3× more reasoning time per minute than GPT-5.3-Codex, which means we’ve gone from ~60 minutes of usable reasoning time per session to just ~20 minutes. That’s not a minor tradeoff — it fundamentally changes how I’m able to use Codex for longer, more complex tasks.
GPT-5.3-Codex struck a great balance between capability and efficiency. Forcing a switch to a model that burns through limits that much faster, without a corresponding increase in allocations, feels like a step backward for users who relied on it.
I’d strongly encourage the team to either restore access to GPT-5.3-Codex, offer an opt-in for users who prefer it, or adjust reasoning time limits to compensate for GPT-5.5’s higher consumption rate.
Thank you for reading — I hope this feedback is taken into account.
5.4 and 5.5 are probably the worst models I’ve ever seen. If OpenAI don’t realize this in time, they will lose thousands of subscriptions to Claude - mine included.
They simply removed your day-to-day model, with no other reliable option: 5.4-mini is just frustating; 5.4 is already too expensive and also dumb, and 5.5 is just extra expensive.
Same here, it really feels like a rug pull, and I’m considering changing to other products subscriptions.
I echo the same sentiments of most of these comments.
5.3 Codex is a really nice fit inside Codex, it’s a clean operator of a model. Whoever tuned the inner details of 5.3 should be celebrated and this should be duplicated for future models running in Codex. My experience with 5.4 & 5.5 is that they are akin to a bull in a china shop on cocaine. Bad focus, jumps around and often makes a huge mess. They are absolutely not “better” when you consider the increased cost. I get the distinct feeling that 5.5s aim is to make extra work for itself not to be helpful but to consume more tokens… (?). Where to from here? Since Codex isn’t near as useful now as it was without 5.3 Codex. The closest and it still burns tokens like a fire starter is 5.5 Low.
It’s all about the dollars in the end..
I have the VS Code extension for Codex. Until yesterday, I was using the Codex model to generate code; however, today I find that I can no longer select the Codex model, but only GPT 5.5, 5.4, and 5.4-mini. My Plus plan is active—the same one I have always had.
Openai did a massive rug pull and is now forcing everyone using the extension to use those three models, completely removed 5.3 Codex and now we all have to use 3x the number of tokens for WORSE results.
The person in charge of making those plans must have changed.
This is common in organizations that begin to decline, often due to personal gain that drives changes to established practices and asserts one’s presence.
Could this be related to partnerships they started with eg. github (copilot) that recently announced gpt-5.3-codex will be a LTS model and their main default? perhaps partnerships with others where the model is still available?
When Codex has more than 5 million weekly active users, and most of these users are pretty happy on a plus plan using an older model (5.3) then what now? Open AI are likely taking a bath on supplying ai compute for this level of 5.3 usage. Adjusting usage limits wouldn’t be cool, so sunset the most in demand model and push everyone to the high usage models. This wouldn’t be so bad if the newer models were that much better to justify the increased price, but I am less productive running 5.4 or 5.5 as it makes a lot of mistakes and requires regular reorientation back to the original prompt goals. And for anyone saying you need to improve / adjust prompting for 5.5. The strength of 5.3 was that it didn’t need to be told what not to do.. I am regularly saying “don’t do that”, and “wtf are you doing that for ffs?”..
I wrote to support, and they told me it was a technical error. Apparently not, unfortunately, but I won’t be renewing my subscription anymore. I need to look for alternatives — I’m sure competitors, especially Chinese specialists, are cleverly taking advantage of this situation. Right now, building agents with competitors is several times cheaper, but it was convenient working with 5.3 Codex. It’s a shame they treat their own users this way.
I enjoyed 5.3 codex model too in the past, but when it was mentioned ages ago that it would be removed, I just looked at what would be a suitable model replacement of 5.3 codex and for me that has been 5.4 mini (mainly) and 5.4 (sometimes).
I have been using those since the announcement on May 27th, just to get in the flow of it and prep myself and honestly those two have been great substitutes. Then I just use 5.5 med for all else and crank up to 5.5 high+ for bigger items.
Really have seen no issues and not sure what all the fuss is about tbh. These things happen. Models change, models leave, models are replaced, the world keeps on spinning. Its just about pivoting and working with what you have and finding a good flow. I also at first though “oh no 5.3 codex is going away”. But I just pivoted, tried 5.4mini and 5.4 to compare against same use cases i did with 5.3 codex, and yeah… smooth sailing tbh.
This is all in the Codex app and not sure how it is (or if it even differs) in the Codex cli.
Excited for 5.6 and 6 to come out soon 
also FYI 5.3 Codex model is still available as a useable model in Cursor. I tried it this morning and it worked fine.
The number of tokens that 5.4 and 5.5 consume. Yesterday I made a simple request for code analysis, and in the end it used 3–4 times more tokens. Working at the same pace, the weekly token limit will run out after 2–3 days of work, whereas before I could work calmly and steadily for the entire week.
What is important for you? People solve their problems or just keep using your modals that generate the money for you, even people already paid for the subscription so they should decide what to use and when.
Its good to improve the modals day by day but to using them users will decide not forcefully.
at lest follow your competitor.
hmm interesting. Just different experiences we are both having then as I have not had the same as you have mentioned at all. So, just different experiences and cases of what is causing this I guess. since using 5.4 mini as my 5.3 repalcement and then doing all else with 5.5 i have not hit the 5 hour or weekly limit at all yet.
I dont think OAI has any bad intentions with removing 5.3. And I dont think its fair to be calling out @VeitB as if they are responsible for this either. they have only been trying to help and be part of this community. I can understand frustration about the model being gone, but I think its also ok to trust the process. I have not had any bad experiences since 5.3 went away, but I respect that this is a major issue for others. I am sure the team is well aware of these frustrations and again i think we should just trust the process and see what they bring next. With the advancements they have been making so quickly already, i dont doubt what the bring next will make users forget about 5.3 not long after.
Compute is limited. They will need to sunset some models to move onto new ones. It’s always been this way.
If they were to stick with 5.3 and never release 5.6 they would fall behind the competition.
But nevertheless feedback is clearly useful because it helps them know when they’ve done something right with a model.