How it works
A small router reads your question, scores every model for it, weighs those scores against price, and calls one model, once.
One question, two steps
- Route. A small router model reads your question once. It writes nothing. For every model on the menu it gives one number: a score for how well that model suits this question. The score is meant to be the chance of a right answer.
- Answer. The model with the best trade between its score and its price answers. One call, and its cost is shown.
In the chat, Preview stops after step one and shows you the ranked list, free. Answer goes on to step two.
The decision rule
Every router follows the same rule. It scores every model and calls the one that comes out highest after a price penalty:
result = the model's score for this question − (price weight × estimated price)
The price weight is the exchange rate between accuracy and money. A higher weight makes the router thriftier; a lower one makes it buy certainty at almost any price. Everything else is measurement.
How a small model can judge a hard question
Judging how hard a question is takes far less than answering it. You do not need to be able to evaluate an integral to recognise one, or to see that "name the capital of France" and "prove this series converges" are different orders of problem. Much of a question's difficulty shows on its surface: its structure, its vocabulary, how many steps it implies.
That is why one pass of a small model is enough, why it writes no text, and why a preview costs nothing to run. It explains why routing is possible. It is not evidence that a given router does it well; that has to be measured.
Why you cannot pick the model
There is no model picker, in the chat or over the API. A picker would hand you the very decision the product exists to make. What you get instead is the full reasoning: every model's score, its estimated price, and which one was chosen. The reply always reports the model that actually answered.
Where the router comes from
Model prices and abilities change every month, so a router that was the best one in the summer is not the best one now. Rather than promise to keep ours current by hand, Fugal runs an open competition on Bittensor: anyone can enter a router, every router is scored on the same new questions each round, and the best one is paid for as long as nothing beats it. See Why a subnet and the Leaderboard.
Right now the chat is serving the router reported live by the service.
A router that comes from the competition is judged there on which model it picks, not on how honest its scores are. So its scores always put the models in order, but they are not guaranteed to be true chances, and a winning router can lean on the kind of question more than on how hard that particular question is. The preview shows such a router's scores as scores, not as percentages. Rewarding honest odds is a change to the competition's rules that is under consideration.
How good it is
The competition is measured in public, every round, and anyone can recompute the results: that is the Leaderboard. No accuracy or savings figure is quoted here for the router the chat serves, because none has yet been measured in a way a reader could re-check.