OpenAI's GPT-5.6 Sol, Terra, Luna hit Bedrock with 90% cheaper prompt caching

Amazon Bedrock has made OpenAI's GPT-5.6 family generally available, bringing three capability tiers to AWS customers. GPT-5.6 Sol targets complex reasoning and agentic coding, GPT-5.6 Terra handles balanced everyday production workloads, and GPT-5.6 Luna is optimized for fast, high-volume tasks like classification and summarization. The big new feature is explicit prompt caching: you choose which parts of your prompt are cached and reused across requests, and cached input is billed at a 90% discount while remaining reusable for 30 minutes. That makes agentic workflows significantly cheaper when system instructions, tool definitions, and reference documents repeat across calls.

The models are served via the OpenAI-compatible Responses API on the Bedrock bedrock-mantle endpoint. Authentication uses short-term bearer tokens generated from your AWS credentials, avoiding long-lived secrets. The model IDs are openai.gpt-5.6-sol, openai.gpt-5.6-terra, and openai.gpt-5.6-luna. Sol is available in US East (N. Virginia) and US East (Ohio); Terra and Luna are also in US West (Oregon). Developers can control reasoning effort from none to xhigh, with medium as default, and GPT-5.6's token efficiency often lets you drop one effort level without losing quality. Streaming, function calling, and strict JSON schema output are all supported.