Prompt Cache Write Costs: How Read-to-Write Ratio Determines Real Savings

AI prompt caching charges a 1.25x write premium but cuts re-read costs to 0.1x the normal input rate, meaning a single cache hit already yields a net saving of 32.5% over two uncached calls. However, analysis of five agent scenarios run through the Synthorai gateway found that net cost impact ranged from an 83% saving in batch workloads to a 6% loss in RAG pipelines. The key variable is the read-to-write token ratio: RAG scenarios produced a ratio of just 0.2 because retrieved documents change with every query, making cache writes nearly impossible to amortize. Claude and GPT-5.6 now share identical explicit caching price structures, while Gemini and older OpenAI models use implicit caching with no write premium but unpredictable hit rates. The practical fix for mixed workloads is selective caching — marking only stable elements like system prompts and tool definitions, while leaving volatile content such as retrieved documents outside cache boundaries.