This could save hours of infra headaches. What’s the latency like when routing across multiple models?