GPT 5.3 Codex removed for ChatGPT subscription accounts despite being listed as supported in the subsciption website

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

• #25830
• #25808
• #25838
• #25781
• #25794

Powered by Codex Action

I'm going to use Claude Code now. GPT-5.3 was the only thing that kept me here at OpenAI.

I was using GPT PRO, but I'm canceling my subscription now; I'm just finishing up this month that's already been paid for. This was a huge lack of respect for users, removing the agent without any prior notice!

The gpt-5.2 and gpt-5.3-codex models have been retired. Please update to a newer model 5.4 or 5.5.

https://x.com/thsottiaux/status/2059650685948551384

The gpt-5.2 and gpt-5.3-codex models have been retired. Please update to a newer model 5.4 or 5.5. https://x.com/thsottiaux/status/2059650685948551384

You should really remove GPT-5.3-Codex from the docs and pricing tables.

<img width="873" height="665" alt="Image" src="https://github.com/user-attachments/assets/3db140d5-1a1a-43b1-94dd-c4dbd5ec5207" />

<img width="1211" height="385" alt="Image" src="https://github.com/user-attachments/assets/c9c387be-d1c2-4eed-9a8c-9e4ed33e3e3a" />

The gpt-5.2 and gpt-5.3-codex models have been retired. Please update to a newer model 5.4 or 5.5.

Thanks for confirming the obvious that the model was intentionally retired.

However, that does not resolve the issue I reported. It actually confirms the core problem: GPT-5.3-Codex was removed for ChatGPT subscription users without clear prior notice, while official OpenAI documentation still presents it as available/usable under ChatGPT subscription plans.
Every GPT model retirement I remember had prior notice weeks in advance, sometimes even a month before the sunset date, usually in the official release notes:
https://help.openai.com/en/articles/6825453-chatgpt-release-notes

But that was not the case here.
Even o3 had its retirement formally announced, but GPT-5.3-Codex did not. You just retired it without formally notifying anyone affected by the change. A single X/Twitter post is not a notice. It is a social media post. Paying customers should not be expected to monitor random social media posts to find out that a model they rely on is being removed.

I should have received proper prior notice weeks ago, either by email or in the official release notes, with the exact retirement date of the model and clear migration guidance. Instead, the model was removed while OpenAI’s own official documentation still listed it as available for paid subscription plans.

That is not acceptable product communication for paying users.
The issue is not simply “which newer model should I use instead.” The issue is that:

1. The official Codex pricing / usage limits page still lists GPT-5.3-Codex under Plus, Pro, and Business usage limits.
2. The same page still includes GPT-5.3-Codex in the credit/cost tables.
3. The ChatGPT release notes did not clearly announce that GPT-5.3-Codex was being retired for ChatGPT subscription accounts.
4. Users were not given a proper migration window or clear in-product deprecation warning.
5. The CLI behavior changed from an explicit error to the model disappearing, while the public documentation still implies the model exists for these plans.

A personal X/Twitter post is not a substitute for official product documentation, release notes, billing/plan documentation, or in-product communication. Paying customers should not have to discover model removals through a third-party social media link after functionality breaks.

Please re-open this issue, or at minimum reclassify it as a documentation/product communication issue rather than closing it as resolved. The following still need to be addressed:

• When will the official Codex pricing and usage-limit pages be updated to remove or clarify GPT-5.3-Codex availability?
• Why was there no official release-note entry or advance notice for ChatGPT subscription users?
• Will there be a temporary compatibility/deprecation period, especially for users who paid for Plus/Business expecting the model listed in the official docs?
• Can the CLI show a clear deprecation message instead of silently removing the model or returning a generic unsupported-model error?

Closing this as “use 5.4 or 5.5” also ignores the practical cost difference between these models.

After the April 9 Codex limit changes, real-world usage appears to be driven largely by reasoning time, not by message count. The practical user model is:

cost of a run ≈ reasoning time × limit percentage consumed per minute

These are not official numbers, but they match the behavior many users have observed in practice, specially me.
On Plus, GPT-5.3-Codex appeared to provide roughly:

• 60 minutes of reasoning per 5-hour limit window
• 60 minutes = 100%
• 30 minutes = 50%
• 1 minute ≈ 1.66%

So a 4-minute task on Plus + GPT-5.3-Codex would cost roughly:

4 × 1.66% = 6.64%

On Plus, GPT-5.4 appears to provide roughly:

• 40 minutes of reasoning per 5-hour limit window
• 40 minutes = 100%
• 20 minutes = 50%
• 1 minute = 2.5%

So a 4 minute task on Plus + GPT-5.4 costs roughly:

4 × 2.5% = 10%

That means GPT-5.4 consumes about:

2.5 / 1.66 ≈ 1.5×

as much limit per reasoning minute as GPT-5.3-Codex.

GPT-5.5 is even worse. In practical use, it appears to consume roughly 2× as much limit as GPT-5.4. That means on Plus:

• GPT-5.5 ≈ 20 minutes of reasoning per 5-hour limit window
• 20 minutes = 100%
• 10 minutes = 50%
• 1 minute = 5%

So a 4-minute task on Plus + GPT-5.5 costs roughly:

4 × 5% = 20%

Compared to GPT-5.3-Codex:

5 / 1.66 ≈ 3×

So users are not being moved to an equivalent replacement. They are being pushed from a model that cost about 1.66% per reasoning minute to one that may cost around 5% per reasoning minute while it spends more time reasoning.

That is roughly 3× more limit consumption even without considering the extra time 5.4 and 5.5 spends reasoning.

The situation is even worse on Business.

Based on user testing, Business appears to provide much less total reasoning allowance than Plus. If Plus + GPT-5.4 gives about 40 minutes per 5-hour window, Business + GPT-5.4 appears to give only about 12.5 minutes.

That means:

• Plus + GPT-5.4: 40 minutes = 100%
• Business + GPT-5.4: 12.5 minutes = 100%

So on Business + GPT-5.4:

• 1 minute = 8%
• 2 minutes = 16%
• 4 minutes = 32%

Compared to Plus + GPT-5.4:

8 / 2.5 = 3.2×

So Business drains about 3.2× faster per reasoning minute than Plus, or about 220% faster.

The reduction in included reasoning time is also severe:

(40 - 12.5) / 40 = 68.75%

So Business appears to provide about 68.75% less practical reasoning time than Plus in this model.

On Business + GPT-5.3-Codex, the estimate was roughly:

• 18.75 minutes of reasoning per 5-hour limit window
• 18.75 minutes = 100%
• 1 minute ≈ 5.33%

So a 4-minute task on Business + GPT-5.3-Codex would cost roughly:

4 × 5.33% = 21.32%

On Business + GPT-5.4:

4 × 8% = 32%

So GPT-5.4 costs about:

8 / 5.33 ≈ 1.5×

as much as GPT-5.3-Codex on Business.

And if GPT-5.5 consumes roughly 2× the limit of GPT-5.4, then on Business + GPT-5.5:

• 6.25 minutes = 100%
• 1 minute = 16%
• 2 minutes = 32%
• 4 minutes = 64%

Compared to Business + GPT-5.3-Codex:

16 / 5.33 ≈ 3×

So a 4-minute task could go from about 21.32% on GPT-5.3-Codex to about 64% on GPT-5.5.

That is not a minor model migration. That is a major practical downgrade in usable subscription value.
GPT-5.3-Codex was the efficient workhorse model: it was capable enough for most coding tasks, tended to spend less time over-reasoning, finished faster, and consumed significantly less of the user’s limit.

Removing it without prior notice forces users onto models that consume substantially more of their paid allowance for the same kind of work.

So, recaping here: this issue is not only about model availability like you said. It is also about:

• misleading documentation;
• lack of formal deprecation notice;
• no clear migration window;
• reduced practical subscription value;
• and a forced move to models that burn through usage limits much faster.

A single X/Twitter post is not proper notice for a paid product change of this scale.

5.3-codex was best working horse, after 2 days coding with 5.4 i'm already get crazy! I spend 2x more tokens to same issues and get nervous breakdown and only half of issues done, 5.4 is skipping half of prompt, forgetting what it doind, stoping and saying: i'm sorry, my fault i should stop here... oh really? The werst thing was, that they even anounce this!! They just remove it.. I tried MiniMax V3 and this this is last month i used chatgpt Pro.

I’ve been a ChatGPT Plus subscriber largely because of Codex 5.3. It was, for me, the best coding model OpenAI has released: fast, reliable, predictable, and efficient enough that the included usage limits were almost always sufficient.

With Codex 5.3 gone, I honestly struggle to see the value of continuing to pay €20/month. I understand the business reasons behind these decisions, but from the outside it feels like another step toward pushing users away from affordable plans and toward more expensive models and higher-priced tiers.

What is particularly frustrating is that many of these newer and more costly models are not necessarily better for practical software development. In many cases, I found Codex 5.3 produced cleaner, more focused, and more useful results than its successors.

I used Codex 5.3 for everything: scripting, debugging, infrastructure work, automation, coding projects, and daily IT operations. It became the tool I relied on most, and the available token budget was generally enough to get real work done without constantly thinking about limits.

What concerns me even more is that this kind of behavior is becoming increasingly common across the AI industry. Companies build trust around a model that users come to depend on, only to later remove it, restrict it, or replace it with something more expensive. This trend should be a wake-up call for the entire industry.

In my view, it is precisely this uncertainty that will push research and adoption toward models that can run efficiently on private hardware. The more dependent users become on centralized providers, the more valuable local, self-hosted, and open solutions become. Democratizing access to AI means giving people the ability to run powerful models on their own infrastructure, without constantly worrying that a provider will change the rules, remove a model, or make their workflow economically unviable overnight.

Now that Codex 5.3 is gone, it feels like the end of an era. To be honest, many of us already suspected this would eventually happen. It's disappointing because, from a user perspective, Codex 5.3 felt like the sweet spot between quality, speed, and cost. OpenAI had something special there, and I think many users will miss it.

decision is simpe as 123 - Unsubscribe

@etraut-openai
Still nothing changed in the website...
This is anti-consumer and misleading, leading customers to believe they will have access to the model when they actually will not.

<img width="992" height="312" alt="Image" src="https://github.com/user-attachments/assets/53beeb82-eeb8-4d43-8404-a6ca5f6ab3a5" />

decision is simpe as 123 - Unsubscribe

probably going to do the same, the value for the price just isn't there. I don't need a vibe coding model or lobotimized mini version of one to have it write the same code as I would, but a little faster. MiMo and Deepseek v4 gave similar results in the free opencode trial and are way cheaper than even the plus subscription for codex.

DeepSeek V4 has been a game changer for me. With roughly €10/month through OpenCode, I can do almost everything I need. The main downside is that inference is currently running in China, which isn’t ideal from a privacy perspective. If DeepSeek V4 becomes available in European datacenters at a reasonable price, I’d happily switch.

That said, I think the bigger trend is local inference. The AI industry should pay attention to this. As models become more efficient and hardware improves, we’re getting closer to running powerful general-purpose models on machines that cost €2,000–3,000 instead of relying on massive datacenters. We still need more optimization and better hardware, whether that’s from NVIDIA, Apple, or someone else, but the direction seems clear.

Today, people struggle to find systems with enough memory, but even DeepSeek V4 appears usable with around 192 GB RAM in some configurations. If, within a few years, high-quality models can run locally at acceptable speeds, many of the assumptions behind the current datacenter-heavy AI business model may start to weaken.

More importantly, this is exactly why the AI ecosystem should avoid becoming dependent on a handful of large providers. The long-term health of the industry depends on pushing toward efficient models that can run on privately owned hardware, making this technology more accessible and more democratic.

I’m currently using Mimo V2.5 Pro, mainly because they lowered the price quite a lot and it feels like the closest thing to GPT-5.3 Codex right now. It is much more direct than DeepSeek V4, which tends to overthink a lot, and from what I’ve seen it has better benchmarks than the V4 Pro version while still being cheaper than GPT-5.3 Codex through the API.

Unfortunately, that is basically what we have left for now. I might try Kimi’s paid plans again, though. Back when they only had K2 and not K2.5 yet, their coding plan was pretty solid, with very high limits and a very good model. I still need to test how it is now after all these months.

The problem with local AI is that there is no real predictability that people will actually have 192 GB of RAM just to run the most basic version of today’s frontier models locally. There will always be a gap between local models and frontier models. The industry has slowed down, and in 10 years the new standard will probably be something like 64 GB RAM (NOT VRAM! So it will be painfully slow), which still won’t be enough to run DeepSeek V4 Flash, which is basically the only frontier-level model that is somewhat cheap in terms of RAM.

In practice, you would need a dedicated AI workstation at home to run models that can actually code well. Even with better quantization, I don’t think local models are going to reach GPT-5.3 Codex level.

We have had a year and a half of progress in tech since o3 was created, and yet no model that runs within 64 GB of RAM, even quantized, has actually surpassed o3. And o3 got outdated a year ago, was not made for agentic coding, and is basically useless for coding today. Gemma 4 32B (no quants) gets close, but it still loses in coding benchmarks.