Thoughts on model routing Model routing is often conceived as the idea that you can optimize cost by routing requests *to the right model* I do not believe this is the right framing, or what routing companies should focus on. Fundamentally, model routing is not a user problem, but a provider problem. AI is expensive, providers want to reduce cost. In the agentic era, workloads are all roughly the same, in the same harness, and there is no better or worse model for a task that isn’t just universally a better or worse model. Caching is such a critical piece to modern inference that the cost of using a small model is negated if the messages later are rerouted to a larger model or vise versa. Anyways it’s not the problem I as a provider care about. The problem I care about is cost, but what I am looking for from the model routers is to get me the best tokens for the price I’m willing to spend. So it’s not routed per request. It’s routed based on the best model I can get, as a provider, for what I can afford to pay. $5/mtok for Opus 4.8? Hell yeah. If it goes above $5, find a GPT-5.5 provider for $3/mtok. If that’s dry, get me GLM under $1/mtok. I don’t want to know why the tokens are that cheap. That’s your job, that’s why I’m paying you to do the routing. Your job is to get the tokens as cheap as you can, through deals or… whatever. Thats what I’d pay for. Most consumer AI is really hamstrung by having to offer pricing that is competitive with OpenAI and Anthropic without being able to subsidize the accounts as heavily. Cost optimization is really the name of the game on pretty razor thin margins.(Shaw (spirit/acc))