Maybe there’s a middle ground where a small local model can roll with the variations in a site that would break a script, while saving the per token costs?