91.2% lower LLM inference cost — and the failed benchmark I almost buried

What building an LLM cost router, finding problems in my own benchmarks, and a corrected 100-query experiment taught me about the infrastructure layer between AI applications and model providers. The