Amazon Bedrock has launched explicit prompt caching for its latest OpenAI GPT-5.6 models, including Sol, Terra, and Luna. These models are now generally available on the platform, offering users the newest generation of OpenAI frontier models with pay-per-token pricing, AWS security controls, and compatibility with existing AWS commitments. The models are divided into three tiers based on their capabilities: GPT-5.6 Sol for complex reasoning and coding tasks, GPT-5.6 Terra for balanced production workloads, and GPT-5.6 Luna for fast, high-volume tasks like classification and summarization. The new feature allows users to control which parts of their prompts are cached and reused across requests, with cached input tokens billed at a 90 percent discount. This is particularly beneficial for agentic workflows where system instructions and reference documents are reused frequently.
Explicit prompt caching on Amazon Bedrock provides users with two modes: implicit, where the platform automatically manages the cache breakpoints, and explicit, where users manually define the cache boundaries for greater control. The feature is designed for write-once, read-many patterns, with cache reads offering significant cost savings compared to uncached input tokens. Users can start with implicit caching, which is enabled by default, and later switch to explicit caching if needed. The platform also supports reasoning effort levels, allowing users to adjust the model's reasoning intensity based on the task's complexity. Additionally, the new models support streaming responses and structured output through the Responses API, enhancing their versatility for various applications.
The GPT-5.6 models are available through the OpenAI-compatible Responses API on the Amazon Bedrock bedrock-mantle endpoint. Users can set up the client using short-term bearer tokens generated from their AWS credentials, ensuring secure and efficient access. The models are available in specific AWS regions, with GPT-5.6 Sol in US East (N. Virginia) and US East (Ohio), and Terra and Luna also available in US West (Oregon). For more details on model availability, users can refer to the Amazon Bedrock documentation.
Source: awsml
