Prompt caching can cut LLM costs by 90%, but many developers are missing it

A developer tested prompt caching across 393 large language models and found that most RAG and agent workloads repeat the same tokens on every request, making caching a major cost factor. Anthropic's Claude requires an explicit cache_control marker to activate caching, meaning teams without it have been paying full price with no warning in the API response. OpenAI's GPT-4o-mini, by contrast, applies caching automatically, delivering roughly 48% savings and faster responses without any code changes. The tests also revealed that inserting even a single varying character — such as a timestamp or session ID — at the start of a prompt breaks the cache entirely, since caching works by matching an exact token prefix. Across the 393 models surveyed, about one-third do not support caching at all, and discount rates range from none to 120x, making cache support a meaningful factor in model selection.