Prompt Caching Cut a 7.3M-Token AI Pipeline Bill by Two-Thirds to $8

A developer running a content pipeline on Claude Sonnet found that a single topic run consumed 7.3 million tokens across 18 AI agent calls, yet cost only $8.12 instead of an estimated $24. The dramatic cost reduction came from prompt caching, where repeated context sent to the model is stored by the provider and re-billed at just 10% of the normal input price. In the benchmark run, 84.5% of all tokens were cache reads, while only 0.2% were billed at full input rates, yielding a cache hit rate of 86.6%. The pipeline — a TypeScript orchestrator producing articles, glossary entries, and hub pages — repeatedly resends the same system prompts and fact sheets with each agent call, making it an ideal use case for caching. The author's key takeaway is that raw token counts are misleading for agentic workflows, and the four-line cost breakdown (input, output, cache read, cache write) is the only figure that truly reflects spending.