Pick a Model, and a Fallback
Configure → Bot Identity → Model Selection lists every model ProxyAI currently offers, grouped High, Mid and Low. Higher tiers reason better on long, ambiguous or multi-step questions; lower tiers answer straightforward ones for less. Beside each name, a green T means the model reads and writes text and a green eye means it also reads images a visitor sends — pick one with the eye if your customers send photos of a product or a receipt.
There is no single right choice, so start from what your bot is for. A bot answering stock, price and opening hours does that just as well on a Low or Mid model, and the saving is real because it is charged on every message. A bot expected to read a long policy document and reason about it, or to hold a complicated complaint together over many turns, is worth a High-tier model. If you are unsure, run a week on Mid and read the transcripts: what you are looking for is answers that were confidently wrong or that lost the thread, not answers that were merely short.
Rates, next to the selector, prices each model before you commit — and every model's cost lands in Usage & costs afterwards, so a switch is easy to judge after the fact. A new bot starts on the cheapest model that still handles storefront questions well. That is a sensible floor, not a recommendation to stay there.
Switching is instant and safe: it changes the next reply only. Nothing about your documents, settings or conversation history is tied to the model you were on.
Set a Fallback Model
Fallback Model sits beside the primary. It steps in only when the primary fails outright — the provider is down, or refuses the request — never because an answer was merely poor. That distinction matters: a fallback is insurance against silence, not a second opinion.
Left on None, a failed reply is retried against the same model and, if that fails too, the visitor gets nothing. Set to a second model, the reply is handed over and the customer gets an answer without knowing anything happened.
A good pairing is a High-tier primary with a Mid-tier fallback: your usual answers stay strong, and the rare outage costs you a slightly simpler reply instead of a dead conversation. Choosing a fallback from a different provider than the primary buys more than choosing a cheaper model from the same one — an outage usually takes a provider down, not one model.
Both models are billed only for what they actually answer, so a fallback that never fires costs nothing to have configured.