Anthropic launches prompt caching: save tokens (and money) when you repeat the same context

Anthropic's prompt caching solves a problem that bleeds money from every AI developer: if you've got a chatbot analyzing a long document (think 50k tokens of technical manual) and your user asks 10 questions, you used to pay 50k tokens × 10 = 500k tokens total.

Now? You pay 50k tokens once, then just the incremental tokens (the question + answer). The savings are wild: Anthropic's claiming 90% less on costs for use cases with repeated context.

How does it work? Claude caches your "system prompt" and long contexts for 5 minutes. Hit the same context again within that window, and you don't repay. After 5 minutes, the cache expires and you're back to paying—but honestly, 5 minutes covers most chat sessions anyway.

Where's this useful? RAG (Retrieval-Augmented Generation) on documents, corporate chatbots running on manuals and policies, agents crawling codebases. Basically anywhere the context is long and you're reusing it.

The catch? It only works on Claude (for now). OpenAI hasn't got anything like it yet, so if you're on GPT you're still bleeding money. But give it time—everyone will copy this.