Sometimes you can get an equivalent result in a fraction of the time using a less capable model.

The problem is of course knowing ahead of time that the faster model can give you the same result.