New: Cookbooks and AI ExplanationsStep-by-Step recipes to solve problems connected to Roadmaps and Cheat Sheets. Need more details? Use AI buttons for structured and simple explanations with concrete examples throughout the whole platform.Take a look
Every request goes to the same expensive model, whether it is a one-line summary or a hard multi-step question.
What you'll have at the end
A router that sends most traffic to a cheap model and only escalates the requests that fail a quality check
You need
An LLM feature already live in production that answers or classifies incoming requests, with at least two models you can call for it: one clearly cheap and fast, one clearly stronger and pricier, and a way to say whether a past answer was right for a handful of real requests.
Not covered
Scoring open-ended free text with no checkable right answer, or predicting difficulty before either model runs; this only fits a task where you can grade the fast tier's own attempt after it answers.
Pin your model's version so a silent update can't surprise you
Go there if the escalation-rate alarm keeps firing because the fast tier's own model quietly changed underneath a name you never pinned.
See your LLM's cost, latency, and errors on one dashboard
Use this first if you have nowhere yet to plot the escalation rate you'll be watching in the last step here.
Checked 15 Aug 2026
Part of the LLMOps cookbook