The cheapest AI request is the one that never reaches a frontier model. Routing gets the least attention of the four cost layers and it's the one I'd ship first. RouteLLM, out of UC Berkeley and LMSYS, reports over 2x cost reduction on benchmarks with no drop in answer quality. In production you can usually beat a trained router, because you already know things a benchmark can't. Prompt caching pays just as fast, and the underrated half isn't the discount. Cache reads don't count against your input-token-per-minute ceiling, so a good breakpoint buys throughput headroom too. Semantic caching I'd scope to a bounded question space and nothing wider. https://www.duskolicanin.com/blog/ai-saas-cost-control-caching-routing-2026 #AI #SaaS #LLM #SoftwareEngineering