> For coding you always want to go with the best model in the category

Will this always be true? There will never be an event horizon/point of diminishing returns where something not-bleeding-edge is "good enough" for 51%+ of users?