Introducing explicit prompt caching for OpenAI GPT-5.6 models on Amazon Bedrock

Amazon Bedrock now offers access to OpenAI GPT-5.6 Sol, Terra, and Luna models, marking a major step in AI infrastructure deployment. These models support a new feature called explicit prompt caching, which allows precise control over which parts of a prompt are cached and reused across requests. This capability significantly reduces inference costs by enabling cached input to be billed at a 90 percent discount, with cached data retained for up to 30 minutes.

The models are accessible via the Responses API, compatible with existing OpenAI SDKs, and support various reasoning effort levels, from none to xhigh, optimizing performance and cost-efficiency. Developers can choose between implicit caching, where Bedrock automatically determines cache boundaries, and explicit caching, where they manually set cache breakpoints for more deterministic control.

Explicit caching is particularly effective in agentic workflows, such as tool-calling loops, where static instructions or references are repeated across multiple requests. By placing cache breakpoints after static content, workloads can minimize redundant computations, lowering costs and improving latency.

The upgrade process for existing users involves changing only the model ID and adopting explicit cache breakpoints, making migration straightforward. The article emphasizes the importance of monitoring cache usage to ensure optimal performance and cost savings.

Overall, this development enhances the efficiency of deploying large language models on cloud infrastructure, especially for enterprise applications involving repetitive prompts or agentic workflows. It also underscores AWS’s commitment to providing flexible, cost-effective AI solutions for developers and organizations.