OutYet reporting
GPT-5.6 Sol on Bedrock makes cache boundaries a deployment decision
AWS's explicit prompt-cache controls for GPT-5.6 put reusable context, rather than only model selection, at the center of cost and latency planning for agent systems.
AWS has documented explicit prompt caching for OpenAI's GPT-5.6 family on Amazon Bedrock, including GPT-5.6 Sol. Instead of relying solely on automatic reuse, an application can identify the end of a stable prompt prefix, give it a cache key, and request explicit caching through the Responses API on Bedrock's bedrock-mantle endpoint. The practical target is repeated agent context: developer instructions, tool definitions, and reference material that stay constant while a user message or the latest tool result changes.
The change is a meaningful departure from the older GPT-5.5-and-earlier behavior documented by AWS, where caching is automatic. OpenAI separately describes GPT-5.6 as adding explicit cache breakpoints and a minimum cache life, while AWS exposes its own request-level controls for the Bedrock route. This is not a new model announcement: GPT-5.6 Sol was already available. The new angle is that teams can now make the intended reusable boundary visible in request construction, which makes cache design part of an agent's integration contract.
The economics explain why that contract matters. AWS and OpenAI both say cache reads receive a 90% discount relative to uncached input, while OpenAI says a cache write is billed at 1.25 times the uncached input rate. Repeated, large prefixes can therefore make explicit boundaries useful for persistent agents, but a one-off request or a prefix that changes frequently has less opportunity to recover the write cost. Technical teams should instrument cached-token and cache-write-token usage instead of assuming that a stable-looking prompt will produce savings.
There are operational constraints alongside the upside. AWS says Sol is available through Bedrock in US East (N. Virginia) and US East (Ohio), and its caching guide says longer-lived cache entries must appear before shorter-lived entries when both are used in one request. AWS also presents caching as a tool for workloads where latency, repeated context, or rate-limit use matter, rather than as a universal performance guarantee. Teams migrating a GPT-5.5 agent should test their actual prompt assembly, region, retention requirements, and cache-hit telemetry before treating the change as a drop-in cost reduction.