Am I using Codex GPT 5.4 wrong?

Codex
1

I am beyond frustrated right now. I need to know if it’s just me or if Codex GPT 5.4 is an absolute regression in terms of coding intelligence.

Disclaimer: I haven’t used older versions extensively for coding, so I’m not doing a historical comparison. I’m just talking about how much 5.4 is ruining my sanity right now.

  • The Infinite Correction Loop from Hell: It feels like I’m babysitting a stubborn intern. I correct Error A, it fixes it but breaks Fact B. I find myself constantly hand-holding and re-correcting the same things over and over. The mental overhead of supervising the AI is starting to outweigh the productivity gains.

  • Encoding Nightmares: It keeps corrupting my file’s character sets. I finish a task only to find all my Chinese comments turned into unreadable gibberish. How is this even possible in a 5.4 model?

  • Flow-Breaking Reconnections: Constant Reconnecting messages. It breaks my concentration every 10 minutes.

  • “Lego-style” Chaos: The architecture it suggests is a fragmented mess. It’s over-encapsulated and functionally chaotic. It’s not “abstraction”—it’s just throwing blocks together without any sense of responsibility boundaries.

  • Pathetic Contextual Association: For a “state-of-the-art” model, the memory is laughable. It ignores global constraints and loses the thread across multiple files almost instantly.

Am I the only one feeling this rage? Is Codex GPT 5.4 just a broken iteration, or am I doing something wrong?

3 Likes
2

Welcome to the forum cg1757834049!

Am I doing something wrong?

In most cases, probably not.

Based on a few years of using generative AI systems, one consistent pattern across widely used public models is performance regression over time—often noticeable after a few weeks. This is neither new nor entirely unexpected, though it remains undesirable.

With experience, you can learn to mitigate these effects to some extent. However, doing so typically requires a more deliberate and defensive approach to how you interact with the system.

The broader perspective still matters. Despite these recurring issues, the overall trajectory remains positive. When comparing the current state of generative AI—with both its benefits and limitations—to where it stood several months or even a year ago, the answer remains consistent: the progress is worth it.

3

No, it’s not broken, but performance is probably very sensitive to the model and level of reasoning you are using - can you confirm you are using the full fat gpt 5.4 with high reasoning?

Can you also confirm the language you are working on and the size of the project - do you have a public repo you are working on?

1 Like
6

I agree, but as a paid product, it is reasonable for users to have higher expectations, especially when it is being compared with other generative AI models.