Has anyone found a sensible way to judge whether an OpenAI model update has actually helped with their day-to-day work? A few outputs may look sharper, but that doesn’t always mean the change matters on real tasks.

I’m considering keeping a small set of recurring prompts and comparing results after updates. What do you include in a useful test set, and how do you account for version-to-version output changes without turning it into a full evaluation project?