Use a cheap model to think. Use a polished one to publish.
The expensive model reads less, so the bill shrinks.
I use reasoning models, at very low cost,
to distill my prompts into smaller prompts.
In my experience they think well but write plainly.
They play chess. They can even build a chess game.
Publishing models, like Claude Fable, render the finished work.
They carry a lot of content and a lot of user refinement.
That polish is why they cost more.
So I let them read the small prompt, not the big one.
We count tokens per query, so the saving shows.
Earthgrid.ai holds 87 models from 7 vendors. We buy from all of them.
I draw the two steps and ask where your tokens go.
Everyone shrinks one real prompt, then we compare.
Two kinds of model, one habit, and a smaller bill.
Tell me the date, the room, and the question you want asked.