Prompt caching & cost basics
Every word (token) you send to and receive from a model costs money and time, roughly proportional to length. Prompt caching is a trick where a long, unchanging part of your prompt (like a big system prompt or document) gets "remembered" cheaply by the provider for repeated calls, instead of being fully reprocessed (and repaid for) every single time.
This matters at the product-building level, not casual chat use - if you're building an app that sends the same 5,000-word system prompt with every single user message, caching that static prefix can cut cost and latency dramatically at volume. Anthropic, OpenAI, and Google all ship this natively in their APIs (usually called "prompt caching" or "context caching") - it's a lever specifically for people building AI products/agents, not something an individual chatting occasionally needs to think about.
No hands-on task needed here - this is a "know it exists" node for when you're evaluating or building a product. Case in point: if you ever price out an AI feature for Kasper's platform, ask whichever provider/dev you're working with whether prompt caching is being used for the static parts of the prompt - it directly affects your running cost.
Who does prompt caching actually matter for?