Amazon Bedrock Adds Prompt Caching for Open-Weight Models
Amazon Bedrock now supports both implicit and explicit prompt caching for open-weight models through the Chat Completions API. Developers can reduce input costs and latency by reusing computed state across requests, with new explicit controls for cache breakpoints.
What changed?
Open-weight models on Amazon Bedrock now support two modes for prompt caching via the Chat Completions API: implicit (automatic) and explicit (developer-controlled). Explicit caching is enabled by supplying the 'prompt_cache_options' parameter with mode set to 'explicit' along with one or more 'prompt_cache_breakpoint' markers in API requests. The cache usage, including tokens used and cache efficiency, is now visible in API response details under 'usage.prompt_tokens_details'.

Why does it matter to an everyday developer?
Many generative AI applications repeatedly resend the same long context—such as system prompts, tools, or reference documents—with each request. Recomputing this context on every call increases both latency and input costs. By leveraging prompt caching, developers can have Bedrock reuse precomputed prompt prefixes across requests, substantially reducing these costs and improving response times. Explicit caching provides fine control over which parts of the prompt are cached and reused, helping developers optimize both application performance and resource usage. However, developers need to consult individual model cards for region and support details, as well as model-specific cache parameters and potential pricing impacts.
What can the developer do now?
- 1
Check Model Support
Verify that your chosen open-weight model and AWS Region support prompt caching modes by consulting the model card documentation.
