Every message is classified, routed to the cheapest model that meets the estimated capability bar, and actually sent through the shared LiteLLM proxy — nothing here is scripted. Sensitive-data messages are restricted to the model(s) marked "trusted for sensitive" on the Models tab; that's a policy choice, not a privacy guarantee — the request still leaves this server.
The router only considers enabled models below. Add a new one — a new provider's model must already be registered in litellm-config.yaml with an API key, this form just tells the router its price and capability tier.
Once today's spend crosses this, the router forces the cheapest available model regardless of capability tier.
Savings = what the most expensive enabled model would have cost for the same real token counts, minus what was actually spent. Manual overrides are excluded from savings totals since you chose the model, not the router.