last baneReported by @dense lotus
#o4 mini and o4-high errors at coding
random marsh
Same to me, o3-mini-high was far better
dusky quiver
The transition to o4-mini-high has drastically reduced code comprehension performance and consistency.
The o3-mini-high model was technically stable, deterministic, and predictable – exactly what I need for productive code use.
Please reinstate this model or provide a functional equivalent.
For production-grade projects, this change is critical.
sleek latch
Same here. This new model even forget code formatting
zealous canopy
same problem. Even custom instructions not working correctly
zealous canopy
I found a perfect solution (as for me), I just switched to platform.openai playground and chose awesome o1-mini-2024-09-12/o1-preview, which was the best models for coding. They don't cut code and fix bugs very carefully and correctly
formal root
The problem started with 4.5GPT that refused to output code more than 400 lines. This contamination has infected all their modals even 4o (march 27th update) and o3 and o4-mini. The modals will get very lazy and not output the full code length no matter the prompting and will omit code or they will use unchanged placeholders and also try’s to condense code like it’s trying to save tokens. The problem gets worse when it decides to use canvas it will just not output the whole code and stop midway through. If u ask it not to use canvas it will then give code that’s not properly formatted with the same issues as without it. I don’t get why they did not add option to disable canvas so it don’t just auto select canvas and even if u do ask it theirs times it will go ahead after few prompts to do it anyway. This issues were not present in o3-mini and o1 and 4o before the updates.
digital mantle
Is there a prompt I can add like "After completing the code, check if the code is complete"? (I tried this, but it didn’t work.)