I typically use models in a learning context, trying to find the edges. It oddly makes sense to use the world's best model and the world's cheapest/fastest model, with little in between. This helps me track the right blend to get things done in practice at a sustainable cost. Today, I feel pretty good about the fable + grok duo.