Bedrock prompt caching can cut repeated input costs by up to 90%

Amazon Bedrock’s prompt caching can reduce input-token charges by up to 90% when an application repeatedly sends the same context, according to AWS pricing cited in a new technical guide. It can also shorten time to the first generated token because the model does not reprocess cached material on every request.

The feature stores a partially processed snapshot of content such as a long document, system instructions or an agent’s tool definitions. A later request can reuse that snapshot while adding a different question. This differs from response caching, which helps only when the entire request is repeated and an identical answer is acceptable.

AWS illustrates the cost problem with a 10,000-token contract paired with 50 questions: without prompt caching, the unchanged contract contributes 500,000 fully billed input tokens. Shortening the prompt or reducing the context window can save tokens, but may remove useful information.

The guide covers six patterns through Bedrock’s Converse API: document content, system prompts, tool schemas, mixed cache lifetimes, tenant isolation and LangChain integration. The advertised savings depend on cache hits, model support, token volume and current pricing. Developers also need to separate tenants correctly and choose cache lifetimes that fit how often shared context changes.