Token anxiety is real. What worked for me: prompt caching on fixed system prompts cut my Anthropic bill by ~60% overnight. Most devs don't realize cache writes are 25x cheaper than input tokens on Claude.

Local models for classification/routing + frontier only for generation is the other move — but the latency tradeoff is real if you're in a user-facing flow.