The teams I see picking cheaper models to save money on coding agents are often the ones spending more. Inference spend is up something like 10x as agents scaled, and downgrading made it worse for a lot of them. A weaker model stalls and retries. Or it creates work a stronger model has to fix later. And switching models mid-session means the new model can't read the warm cache, so you're paying full price to re-read context you'd already paid for. The cheapest call is rarely the cheapest way…