I find using different core models more useful than just picking one "best" one.
It's like getting different perspectives. They all find things the others miss and approach a problem differently.
Models from the same family or with different prompts are too similar.