The gateway already load-balanced across providers on a single key. Now you can set a confidence condition, and when the primary model's answer trips it, the request escalates to a fallback you've named. That's useful if you route most traffic to a cheap model by default and only want to pay for the big one when the small one starts hedging. The same October 6 update widened the roster reachable through that one key: Mistral Large 4 landed, Google's Nano Banana 2.1 image model showed up, and Microsoft's latest speech and transcription models came online. The model additions are routine. The confidence hook is the piece worth wiring in.