Cached input tokens are cheaper because the GPU avoids recomputing the same intermediate state (KV cache) for repeated conversation prefixes.
Key-value caching (KV caching) stores the model's internal state after processing initial messages, so subsequent calls with the same prefix skip redundant computation.
Providers pass these compute savings to users as lower per-token prices for cache hits compared to cache misses.
To maximize cache hits, avoid changing inference settings or previous messages between prompts in an application.