Amazon Bedrock now supports explicit prompt caching for open-weight models, allowing applications to reuse the computed state of long instructions, reference documents and tool definitions across requests. Matching content is read from cache instead of being processed again, which can reduce input charges and response time.
Developers can use implicit caching, which automatically looks for eligible repeated prefixes, or explicit caching, which places breakpoints after content that should be reused. Explicit mode is useful when a stable shared prompt is followed by a question that changes on every request, because it avoids paying to write one-off content into the cache. Several breakpoints can represent layers that change at different rates.
Support varies by model and API. Each model can set its own minimum prefix length, number of breakpoints and cache lifetime. Cache reads cost less than ordinary input, but some models charge more for the initial cache write, so savings depend on how often the same prefix is requested. Keeping static content first and dynamic material last improves the chance of a match.
The feature works through Bedrock’s Chat Completions API and returns cache-usage fields with each response. AWS says prompts and outputs are not shared with model providers or used for training, while access, model licensing and regional availability remain separate requirements.