What the router sees

A router that only reads the request text can only guess. Emissary's router also sees the situation around the request, and that's what lets it find the cheapest good option rather than the cheapest nominal one. This page covers each signal. For how they combine into a decision, see how a routing decision is made.

Cache state

Also called: stateful routing.

Prompt caching changes the economics of a request. Sending a long conversation to a model that already holds it in cache can be far cheaper than sending it to a nominally cheaper model that has to process it from scratch. A router that can't see cache state can't make that tradeoff.

Emissary's router knows what each model has cached and takes it into account. Each target model's cache read and write prices are part of its configuration, so cached context is priced in.

Expected output length

Also called: intelligence density.

Models differ in how many output tokens they need to complete the same task. A model with a lower per-token price that needs far more tokens to answer can cost more overall. The router estimates how many output tokens each model will need for the task and prices that in, instead of comparing per-token prices alone.

Effort

Emissary routes across models and effort, not just models. Many models can run at different effort levels, trading more reasoning for more cost and latency. For a request that doesn't need a different model, the router can change the effort instead.

The classifier has one head per model and effort, for example @low, @medium, and @high for a model that supports them. Each head is the probability that the model passes the request at that effort. A model is confident when any effort clears the confidence threshold, and it's served at the lowest effort that does. Models without effort levels have a single head and run at the provider's default effort.

This matters because it preserves the cache. Prompt cache lives with the model, so switching models forfeits whatever the previous model had cached. Staying on the same model and changing the effort keeps it warm. A router that can see cache state prefers the cheaper move: adjust effort first, and switch models only when the capability gain is worth losing the cache.

Where effort routing applies to a model, the Effort routing column in the router's Configuration tab shows it.

Modality

A router needs to know which models can accept which inputs. The router knows which models handle images, shown in the Vision column of the Configuration tab, so it never sends an image request to a text-only model. Being modality-aware doesn't mean the router interprets the image content. That is a further step, called multimodal routing. See the design space.

Your pricing

The ranking uses the prices on your router: input, output, and cache read and write, per model. Pricing is prefilled with list prices, and you can change it to match your contracts. That way the router's idea of "cheapest" is your cheapest. See use your own pricing.

Context limits

Each model's context window is part of its configuration, so a request that's too long for a model isn't routed there.

At a glance

Signal	What it prevents
Cache state	Paying to re-process context another model already holds
Expected output length	Choosing a "cheap" model that costs more once it answers
Effort	Losing the cache by switching models when changing effort would do
Modality	Sending an image to a model that can't read it
Your pricing	Optimizing for list prices you don't actually pay