I cannot replicate the problem.
Lets make the first turn of a chatbot in the playground:
Pound on the “view code” button in the playground after constructing your input and settings. You then only need one line added to get the clean output:
The very low setting of top_p ensures not just similar quality, but near-identical output if that’s what you need to verify.
(We get the objectionable “Ah!” exclamation watermark of the AI not believing in leprechauns)
Azure models of similar or the same name aren’t necessarily in sync with the internal versions that OpenAI runs.