What is Prompt caching?
Prompt caching reuses a model's computation for repeated prompt prefixes, billing the cached input at a fraction of the normal rate.
Many applications send the same long prefix on every request — a system prompt, a tool schema, a document, or a conversation history. Prompt caching stores the model's internal representation of that prefix so it doesn't have to be recomputed on the next call, which is both faster and cheaper.
How much cheaper varies by model, and by more than most pricing pages admit: across the models Kyma publishes a cached rate for, the cheapest bills a cached token at a tenth of the input rate and the dearest bills it at the full input rate — no discount at all. Some models publish no cached rate at all. For agents and chatbots with large stable prefixes this is often the single biggest lever on cost, which is exactly why it is worth reading the model's own figure rather than a headline percentage.
Caching is most effective when the cacheable prefix is large and stable and the variable part (the user's new message) is small — exactly the shape of coding agents, RAG pipelines, and long-running chats.
How Kyma does it
Kyma forwards caching hints to models that support it and publishes each model's own cached rate beside its input rate on /models and /pricing — a dash where nobody publishes one. Every response reports usage.cost and the cache discount, so what caching saved on your request is a measured number rather than an advertised one.