OpenAI GPT-5.6 models launch on Amazon Bedrock with prompt caching

OpenAI's GPT-5.6 launch on Amazon Bedrock extends the frontier model family to AWS's managed service platform, giving customers pay-per-token pricing alongside AWS security and governance controls. The introduction of explicit prompt caching addresses a specific pain point in agentic workflows, where system instructions, tool definitions, and reference documents remain static while user queries change on every call. By allowing developers to mark exactly where the reusable prefix ends, explicit caching offers more predictable hit rates than implicit caching, which Bedrock automates on its own. The pricing structure—90 percent discount on reads, 1.25× on writes—is designed to reward write-once, read-many access patterns typical of agent loops that call the same model repeatedly within a 30-minute window.

The three-tier model hierarchy reflects OpenAI's strategy of letting customers choose reasoning intensity per workload. GPT-5.6 also supports adjustable reasoning effort levels (none, low, medium, high, xhigh), allowing operators to balance latency and quality; the body notes that GPT-5.6 is more token-efficient than earlier versions, so many workloads can maintain quality at a lower effort setting than they used on GPT-5.5 or GPT-5.4. Availability varies by region—Sol in two regions, Terra and Luna in three—constraining initial adoption to US deployments.