Cache Read
One of the billing items for large model APIs, referring to the portion of a request that hits the server-side cache and is billed at a discounted unit price, typically one-tenth of the standard input price. It is the largest cost component in agent scenarios.
Timeline (last 90 days)
Autonomously running AI repeatedly references the same contextual base, achieving cache hit rates of over 80%, which is why agentic tasks see greater cost reductions.
Under the new pricing, the cache read price dropped 75%, to just $0.25 per million tokens.
Cache Read refers to the model directly reading intermediate computation results of already-parsed context from cache during multi-turn conversations or repeated calls, without needing to reprocess.
All Facts (2)
Autonomously running AI repeatedly references the same contextual base, achieving cache hit rates of over 80%, which is why agentic tasks see greater cost reductions.
50%UnverifiedCache Read refers to the model directly reading intermediate computation results of already-parsed context from cache during multi-turn conversations or repeated calls, without needing to reprocess.
50%