GPT stealing model is disgusting, $200 is wasted

I completely agree with your experience. My usage quality has significantly deteriorated since August 28, 2025. Before that date, both the main application and Codex were really satisfying and had clear strengths compared to Claude Code. But now the situation has become unacceptable.
As a PRO subscriber paying premium prices, I'm experiencing constant issues even with basic functionality. The model switching you described is just one part of the problem - there are errors in elementary tasks that shouldn't happen at this tier of service.
This dismissive attitude towards paying PRO users is extremely disappointing. We're essentially funding the service while receiving degraded performance. If nothing changes, I'm planning not to renew my subscription when it expires. For $200/month, we deserve transparency about model usage and consistent high-quality performance, not this bait-and-switch approach.
It's frustrating because the platform had real potential and unique advantages, but the recent quality drop and lack of acknowledgment from the OpenAI makes it hard to justify continuing with the service.

我完全同意你的感受。自2025年8月28日以来，我的使用体验显著下降。在此之前，主程序和Codex都非常令人满意，并且与Claude Code相比优势明显。但现在的情况已经令人无法接受。 作为一名付费的PRO用户，即使是基本功能也经常出现问题。你所描述的模型切换只是问题的一部分——一些基本任务中出现了一些错误，而这些错误在这个级别的服务中不应该发生。 这种对付费PRO用户的轻视态度令人非常失望。我们实际上是在为这项服务付费，却只能承受性能的下降。如果没有任何改变，我打算在订阅到期后不再续订。每月200美元的费用，我们应该获得模型使用情况的透明度和始终如一的高质量性能，而不是这种“诱饵换货”式的收费方式。 这令人沮丧，因为该平台确实有潜力和独特的优势，但最近质量的下降以及团队缺乏认可，使得我们很难有理由继续使用该服务。

Yes, I just paid $200 for Pro, and since I just opened Plus, I can't upgrade directly to Pro, which is a waste of $20. But GPT5 High is indeed better than Claude overall. Claude isn't that smart, and many problems can't be solved well, but GPT5 helped me analyze and solve them. It's a pity that their strategic risk control is so poor. I will go back to Claude after it expires next month. I'm waiting for Google 3.0 to try it out. Google 2.5 Pro's coding ability is a bit weak, and it's even more deceptive and boastful than Claude.

> 我完全同意你的感受。自2025年8月28日以来，我的使用体验显著下降。在此之前，主程序和Codex都非常令人满意，并且与Claude Code相比优势明显。但现在的情况已经令人无法接受。 作为一名付费的PRO用户，即使是基本功能也经常出现问题。你所描述的模型切换只是问题的一部分——一些基本任务中出现了一些错误，而这些错误在这个级别的服务中不应该发生。 这种对付费PRO用户的轻视态度令人非常失望。我们实际上是在为这项服务付费，却只能承受性能的下降。如果没有任何改变，我打算在订阅到期后不再续订。每月200美元的费用，我们应该获得模型使用情况的透明度和始终如一的高质量性能，而不是这种“诱饵换货”式的收费方式。 这令人沮丧，因为该平台确实有潜力和独特的优势，但最近质量的下降以及团队缺乏认可，使得我们很难有理由继续使用该服务。 Yes, I just paid $200 for Pro, and since I just opened Plus, I can't upgrade directly to Pro, which is a waste of $20. But GPT5 High is indeed better than Claude overall. Claude isn't that smart, and many problems can't be solved well, but GPT5 helped me analyze and solve them. It's a pity that their strategic risk control is so poor. I will go back to Claude after it expires next month. I'm waiting for Google 3.0 to try it out. Google 2.5 Pro's coding ability is a bit weak, and it's even more deceptive and boastful than Claude.

Funny coincidence - I had the exact same Plus/Pro upgrade issue! Was paying through Apple and had to contact Apple support to cancel Plus first before I could upgrade to Pro.

> > 我同意你的感受。自2025年8月28日以来，我的完全使用体验显着下降。在此之前，主程序和Codex都非常令人满意，并且与Claude Code相比有明显优势。但现在的情况已经无法接受。即使作为付费的PRO用户，基本功能也经常出现问题。你所描述的模型切换问题只是部分——一些基本任务中出现了一些错误，而这些错误在这个级别的服务中不应该发生。这种对付费PRO用户的轻视态度令人非常失望。我们实际上是在为这个服务付费，却只能承受性能的下降。如果没有任何改变，我打算在订阅后不再续订。每月200美元的费用，我们获得模型使用情况的间隔并且始终如一的高精度成绩，换成这种“诱饵货”式的收费方式。这令人欣慰，因为该平台确实有潜力和独特的优势，但最近质量的下降以及团队缺乏认可，使我们很难有理由继续使用该服务。 > > > 是的，我刚花了200美元买了Pro，而且因为刚开通了Plus，所以不能直接升级到Pro，浪费了20美元。不过GPT5 High确实整体比Claude好用。Claude本身不太聪明，很多问题解决不好，但GPT5帮我分析解决了。可惜他们的策略风控太差了。下个月Claude到期后我会回去用，等谷歌3.0试试。谷歌2.5 Pro的编码能力有点弱，比Claude更坑爹，更吹牛。 巧合的是——我升级 Plus/Pro 时也遇到了同样的问题！之前是通过苹果支付的，所以必须先联系苹果客服取消 Plus 才能升级到 Pro。

I don't know what to say anymore. I spent $200 and it's totally useless. What am I using gpt4o 4.1 for? It's worse than Claude 3.5. I can only contact the official customer service to see if there is any whitelist.

I paid and selected the actual model, GPT5 High. However, the actual output execution often alternates between GPT4.1 and GPT4o. Sometimes, I even ask them to analyze the solution using GPT4.1.

How does one tell of the output is from GPT 4o instead? I've had a great experience so far (using GPT 5 high, Pro version, switching from Claude Code) and have had no issues.

> 我付费选择了实际使用的模型 GPT5 High。然而，实际的输出执行经常在 GPT4.1 和 GPT4o 之间交替进行。有时，我甚至会要求他们使用 GPT4.1 来分析解决方案。 如何判断输出结果来自 GPT 4o？到目前为止，我的体验非常好（使用 GPT 5 high，专业版，从 Claude Code 切换过来），没有遇到任何问题。

Before answering each question, mark the current model in the reply. Write a small label on the first line: [Model: xxx] to facilitate user confirmation

The only supported model for Codex is GPT-5. We don't change the reasoning effort nor the model at anytime.

Why do you think the model changed to 4o?

As mentioned above, codex uses the "gpt5-codex" model, which is a version of gpt5 tuned for code generation tasks. It doesn't fall back on 4o or older reasoning models.

I'm going to close this issue because I don't think there's anything actionable here, but if you have any questions about the models used in codex, feel free to post in the discussion section. And if you're seeing a behavior that you think may be a bug, please file a new bug report with repro details.