Prompt Caching (Cached Tokens)
Reusing a previously processed prompt prefix at a steep discount — often 90% off the input rate on a cache hit.
Prompt caching lets you reuse the model’s work on a repeated prompt prefix instead of paying full price to reprocess it every call. When a request reuses cached content — the same long system prompt, document, or tool set as a recent call — those cached input tokens are billed at the cached-input rate, which is dramatically lower than the standard input rate.
The discount is large: several providers charge one-tenth of the input rate for a cache hit, and DeepSeek charges a small fraction of a cent. This makes caching the single biggest lever for agents, chatbots, and RAG systems that resend a large, stable prefix on every turn. This explorer shows the cached-input rate as its own column wherever the provider publishes it; a dash means the provider does not list one.
Caching is not free to set up. Some providers charge a one-time cache-write fee to store the prefix, and cache entries expire after a few minutes unless refreshed. The savings only materialize when your hit rate is high — if the cached prefix changes every call, you pay the write cost without the read discount. Model the % cached field in the hub calculator to see the effect on your bill.
FAQ
How much does prompt caching save?
A cache hit is typically billed at about 10% of the standard input rate, and even less on some providers — so a heavily cached workload can cut input cost by roughly 90%.
When should I use prompt caching?
When you resend a large, stable prompt prefix — a long system message, document, or tool set — across many calls. High reuse is what turns the cache-read discount into real savings.