LLM routing, explained

LLM routing is choosing, for each request, which model and which provider should answer it. A router first keeps only the models that can do what the request needs, then picks among them by price, speed and health, so simple requests go to cheap models and hard ones to strong models.

Start free Read the docs

Asked forauto (balanced)

Needstools, 2,100 tokens in

Shortlist, per 1M tokens in / out

claude-haiku-5-5Anthropic · $0.10 / $0.50
gpt-6-lunaOpenAI · $0.10 / $0.50
gemini-3.8-flashGoogle · $0.75 / $3.75

Chosen: claude-haiku-5-5 on Anthropic, first attempt

Two kinds of routing

Kind	Chooses	Example
Provider routing	Which provider serves a named model	The same model from its maker or a cloud host, whichever is healthy and cheaper
Model routing	Which model answers at all	"auto" sends a short summary to a small model and a hard refactor to a large one

An engraving of a wall-mounted telephone transmitter from a 1906 telephone handbook. — From a 1906 handbook on telephone exchanges.
The practical telephone handbook and guide to the telephonic exchange, 1906. Internet Archive Book Images, no known copyright restrictions, via Wikimedia Commons.

How a router decides

Read the request.
Does it include images, tools, a JSON schema, a long context, a reasoning setting?
Filter.
Drop every model that can't do all of it, or that your rules forbid.
Order.
Healthy providers first, then by your strategy: lowest price, lowest latency, or a balance.
Keep a fallback list.
The next choices are kept in case the first one fails.

Why it saves money

Model prices differ by around a hundred times: on DigitSpace today the cheapest text models cost about $0.10 per million input tokens and the most expensive about $10. If most of your traffic is simple, sending it to a small model instead of a flagship cuts that part of the bill by an order of magnitude, while hard requests still get a strong model.

Questions

A telephone exchange room around 1900, with a long wooden switchboard, operator stools and a desk. — General Telephone Company of the Northwest, exchange room, about 1900, public domain, cropped, via Wikimedia Commons.

Isn't a cheaper model worse?

For hard tasks, often. Routing with a quality floor only picks among models good enough for the level you set, and named models are never swapped.

Does routing need an extra AI call?

Not on DigitSpace: the decision uses the request's features and live price and health data, without asking another model.

Live in a minute.

Create a key, change the base URL, send a request. No card needed with your own provider keys.

Start free