Tutorial: routing LLM requests to the cheapest capable model

On November 4, 2026, a post on the DEV Community AI channel titled Saving Money with LLM Routing: Sending Simple Requests to Cheap Models opened with a cautionary example. Someone asked whether the Globex contract could be ended early. The question was seven words long, looked short and simple, so the router sent it to the cheapest model. The answer came back about half a second later: yes, just give 30 days' notice. The contract in fact called for 12 months plus a penalty, so the answer was wrong, and the router itself had picked the model that produced it. The post argued that routing can cut a large share of an LLM bill, but not in a way that trades correctness away. The approach it put forward was to assign each request to the cheapest model capable of handling it, and to add a check that catches the mistake when routing does go wrong, so that cost savings do not come at the price of an incorrect answer. The article said it would build such a router and describe the check that catches these failures.