The five model providers ResponseCX routes to, the default model per provider, and how to pick per agent.
ResponseCX is provider-agnostic: every agent carries a model field, and the platform routes the
request to the matching provider. Leave it unset and the provider’s default applies.
Any model identifier the provider accepts is passed through, so a newer model is available the day
the provider ships it — no platform release needed.
Match the model to the step, not the agent fleet. A common shape: a fast, cheap model
(llama-3.1-8b-instant or gemini-2.0-flash) for intent classification, and a stronger model for
the customer-facing reply. The classification step runs on every message; the strong model only on
the ones that need it.
Model choice affects fluency, latency, and cost. It does not decide what the agent is allowed to
do — that is Settings, Rules,
and, for consequential actions, the decision gate.
Upgrading a model is a behaviour change. A stronger model follows your policies differently — often
better, occasionally in a new wrong way. Run your
evaluations against the new model before switching production traffic.
