Cached input pricing
A discount on input tokens that repeat across calls - a stable system prompt, few-shot examples, a long document. The cached prefix is billed at roughly 10-25% of the normal input rate.
What it is
When consecutive requests share an identical opening prefix, the provider can reuse the computation for it. You pay a small one-time "cache write" fee, then a heavily discounted "cache read" rate on every later call that hits the same prefix (within a short TTL).
It is a billing feature, distinct from the in-GPU kv cache.
Why it matters
For any app with a large fixed prompt - agent instructions, a schema, a knowledge base - caching can cut the input tokens bill by 50-90%. It only works if the prefix is byte-identical and requests are frequent enough to stay warm.
Example
A coding agent with a 12K-token system prompt, 100K calls/month: uncached input is ~1.2B tokens. At $3/1M that is $3,600. With cache reads at $0.30/1M the same prefix costs ~$360 - a $3,240 saving.