Do we really need the most powerful LLM for every request? Probably not. One thing I’ve been thinking about while working with AI systems is that model selection itself is becoming an engineering problem. Not every request needs the same model. For example: → A simple classification or extraction task can probably run on a smaller, faster model. → A complex reasoning problem may justify a more capable model. → An image/audio request needs a model designed for that modality. → An agentic workflow may need a completely different setup again. So instead of sending everything to one expensive model, we can put a routing layer in front of multiple models. Something like: Request → Router → Choose Model → Response The router can consider things like: Task complexity Expected quality Latency requirements Token usage Model cost Modality Previous conversation/context This makes LLM infrastructure less about "Which model is the best?" and more about: "Which model is good enough for this particular request?" That distinction can have a pretty big impact on cost, latency and scalability when you're operating an AI system at scale. I recently came across LLMRouter, an open-source project working on this exact problem, with multiple routing strategies for single-turn, multi-turn, multimodal and agentic use cases. Interesting direction for anyone building multi-model AI systems. Would you use an LLM router in a production architecture, or would you keep model selection explicit at the application layer?