+
Definition
+
+ This is the average token cost incurred by the router's selected models for 1000 queries from our dataset.
+
+ We obtain the per-token cost for the specific models a router
+chooses using the official API pricing published by their providers. For unpopular models that are not served by commercial providers, we deploy them ourselves for experiments.
+In such cases, we approximate their costs using the pricing tiers published by commercial hosting
+platforms.
+
+
+