LLM API costs are the fastest-growing unmanaged line item in enterprise IT, and most security teams have zero visibility into how models are selected, what data enters context windows, or what the real bill will be until it arrives. This talk breaks down the cost mechanics of multi-turn AI deployments across OpenAI, Anthropic, and Google, shows how prompt caching and context engineering reduce costs by 50-90%, and makes the case that AI cost governance belongs on the CISO's desk, not buried in a DevOps backlog.