What is an AI model router?

An AI model router is the part of a system that decides which model answers each message. You write one thing; something has to choose whether a small fast model or a large expensive one deals with it. In most chat apps that choice does not exist, because the app only has one model family to offer. A router is what you build when you have many.

The reason routing matters is that models are not interchangeable, and neither are their prices. Across the current frontier, the gap between the cheapest capable model and the most expensive one is ten to fifty times, per token. The strong models are genuinely better at hard reasoning, long documents and careful judgment. They are not meaningfully better at telling you the capital of Australia. Sending every message to a flagship is paying limousine rates for grocery runs. Sending every message to the cheap model saves money until the day it quietly gives you a wrong answer that mattered.

What a good router weighs

A useful way to think about a router is as triage. For every message it asks a few questions. What does this need, technically: does it include an image, a huge document, a request for code? How hard is it, honestly: a lookup, a routine draft, or something that rewards real thought? What has the conversation been doing, since switching models mid-thread has costs of its own? And what does each candidate cost right now? The router's job is the cheapest model that will genuinely nail the message, where genuinely is doing all the work.

The trust problem

Routing has an obvious failure mode: the company saves money by quietly giving you the cheap model, and you can never tell. This is why we think a router owes you two things. First, visibility: in Heuricity every routed message carries a small label naming the model that answered and the reason, like quick lookup or deep reasoning. The judgment is printed on the page, not hidden in a server. Second, a quality bar you could check: we test routed answers blind against always-using-the-flagship, and a routed answer that a reader can distinguish as worse counts as a routing failure, not a saving.

Get those two things right and routing stops being a cost trick. It becomes the feature: the reason one subscription can cover every frontier model and still cost less than any single one of them.