i have had similar experiences. o3 did a great job not a few weeks ago dissecting large complicated code bases, refactoring complex components.
Now the last couple of weeks it is utterly useless.
Ignores instructions continuously and without exception
Hastily concludes
Invents information
It is unusable. Other models that i have access to such as GPT4.1 or similar are just not advanced enough to be useful in a coding context
From what I have observed, this is the beginning of the downfall of openai.
Openai had a lead, and openai sat on that for far too long.
Instead of attempting to maintain that lead, openai tried to expand, while publicly stating that GPT 5 was delayed. Even stating that GPT 5 is simply a unification of all models.
I have yet to see a script produced by chatgpt that did not instantly crash. This includes scripts o1-mini wrote. On rare occasions, I will see an output that far exceeds what is normal for these models. This only occurs when both my usage has been minimal for a week, and the servers are at low capacity. Essentially, to save money, quantization is being used. If that was not enough, these models are setting a maximum output token limit. If you give the model a 250 line script, you get one that is 150 lines. This is due to what is in the system prompt that o4-mini shared with me at random one day.
Each of the other 3 competitors saw this, and took advantage of it. In reality, the only thing holding them back was openai.
Claude shifted to vibe coding. Claude’s free model can produce better scripts than openai’s, every single time.
Grok has gone for the everyday user. Grok released actual avatars for its model, taking whatever was left.
Gemini is what chatgpt used to be. Google pushed for intelligence, far exceeding openai’s models with a wide lead on understanding and intelligence. Also released better video generation than sora. This is who wins the ai race. Google is not dependent, like openai. Google has tpus, openai relies on nivida for gpus. Google does not need Gemini for income.
The way I always explain it, is that Gemini is working on a project with 2500 lines of cpp while chatgpt cant even output python scripts. Gemini has been working on a project that nobody has ever made before, and chatpgt cant provide a script for ffmpeg. Gemini is has been writing cpp from scratch, chatgpt cant refactor a 250 line script that it made in the past.