One frustration I've had this week is some LLM versions work much better and worse for some prompts. For example, some (even in same family or same version) consistently had a parse error on things that the rest of the LLMs could handle. I was irked, but shrugged it off.