Key highlights in addition to the model quality itself:

* real-time router that quickly decides which model to use based on conversation type, complexity, tool needs, and explicit intent (for example, if you say “think hard about this” in the prompt)

* router is continuously trained on real signals, including when users switch models, preference rates for responses, and measured correctness, improving over time.