Inference Cost Levers: What to Cut First, What It Saves and When It Backfires

Inference costs come down through seven main levers: removing unnecessary work, trimming context, prompt caching, output limits, routing by difficulty, batch processing, and quantization or self-hosting. This guide orders them by risk to quality, gives the evidence for each, and names the condition under which each one fails.