@emollick I spend a lot of time on this problem. I work on Auto style routing across Microsoft/GitHub and maintain vLLM Semantic Router. The key distinction I'd make from the Jagged Frontier framing is: human perceived difficulty != model capability != router quality Good routers are not just doing "easy task to small model, hard task to big model." They are trying to predict which model is best for a specific request using signals like task type, capability profile, context, cost, latency, confidence, and failure modes. That's the approach we're taking with Auto. Models have different capability surfaces. A smaller model can beat a larger one on some tasks, while the reverse is true elsewhere. So routing is really a model selection problem, not just a difficulty estimation problem. There's also a growing body of empirical work showing that routing can outperform fixed model setups on quality, cost, or both. We summarize that literature and point to the relevant results in our paper: arxiv.org/abs/2603.21354 and arxiv.org/abs/2603.04444 To me, the frontier is how accurately we can learn: request -> best model Over the past year, vLLM Semantic Router has been exploring this from several angles: signal driven routing, context aware model pools, learned model selection, routing performance, and policy/runtime integration. I'm much more optimistic about Auto.