Heterogeneous routing and the cost of thinking

Not every request deserves the same amount of thought. Deciding how much is its own system.

Give every task the largest model and routine work inherits an expensive habit. Give every task the cheapest model and difficult work costs you in retries, corrections, and failure.

Routing has a cost. So do checking, retrying, and recovering from a wrong answer. Count all of them.

Cheap tokens are easy to price. Finished work demands harder accounting.

Unobe Corp