I’m considering using an OpenAI-compatible SDK to make requests to DeepSeek, but I’m not sure how far that compatibility goes once I use a reasoning model. I’d rather find the rough edges before building a client around assumptions that only hold for standard chat completions.

Which request parameters, streaming behaviors, or response fields have you had to change when switching between providers? I’m particularly curious about gotchas that don’t show up in a basic “send a prompt, get text back” test.