How should Model selection and routing be designed for production?

Benchmark candidate models on real workloads, define routing rules, use fallback behavior, monitor quality by route, and keep provider-specific details behind an abstraction when useful.