GPT-5.6 Sol on Bedrock adds a second production path, with region and caching tradeoffs

AWS now offers GPT-5.6 Sol through Amazon Bedrock's Responses API, with the Sol tier listed for US East (N. Virginia) and US East (Ohio). AWS says pricing matches OpenAI's first-party rates and that usage can count toward existing AWS commitments. This is a platform-access update rather than an OutYet release determination: availability reporting does not change model status, which remains the responsibility of provider detectors.

The timing matters because OpenAI's July 9 product announcement framed Sol as the flagship reasoning tier in a three-model GPT-5.6 family, alongside Terra and Luna. OpenAI says the new tier names are intended to represent durable capability levels, while AWS presents the family as an option for customers that need AWS-region deployment, IAM controls, VPC use, and CloudTrail logging. Those are different layers of the product: model capability comes from OpenAI, while the service boundary and operational controls come from Bedrock.

OpenAI reports that Sol reached 73.5% on ExploitBench 2 versus GPT-5.5's 47.9% at a comparable output-token budget, and that Sol reached 80 on the Artificial Analysis Coding Agent Index with maximum reasoning. AWS repeats the coding and cybersecurity positioning, but neither official page substitutes for a workload-specific evaluation. Teams comparing Sol with GPT-5.5 or another frontier model should therefore treat the reported figures as useful vendor-supplied evidence, then test their own tool loops, prompts, latency targets, and failure cases.

The Bedrock-specific technical change is prompt caching with explicit cache breakpoints. AWS says reusable prompt context can be retained for at least 30 minutes and cached input receives a 90% billing discount, which can matter for agent runs that repeatedly send the same system instructions, tool definitions, or reference material. The benefit is conditional: a workflow that changes most of its context on every call will have less reusable input, so cache design becomes part of the application architecture rather than an automatic cost reduction.

There are practical limits to the new route. AWS lists Sol in two regions, while Terra and Luna have a third US West region, so deployment locality may decide whether Sol is viable for a particular workload. OpenAI's first-party announcement also describes reasoning settings and multi-agent capabilities, but the AWS announcement focuses on Bedrock access through the Responses API. Engineers should verify the selected region and current API support during implementation instead of assuming that every first-party capability maps identically to every hosted configuration.