Amazon Bedrock adds a hosted route for GPT-5.6 Sol, with migration and governance trade-offs

AWS says GPT-5.6 Sol, Terra, and Luna are now available through Amazon Bedrock, using the OpenAI Responses API on its bedrock-mantle endpoint. The practical change is a hosted deployment route for an existing model family: Sol is positioned for autonomous coding, security research, scientific analysis, and deeper multi-step reasoning, while Terra and Luna target progressively more cost- and latency-sensitive work. AWS lists Sol in US East (N. Virginia) and US East (Ohio), whereas Terra and Luna also appear in US West (Oregon).

The timing matters because OpenAI's own GPT-5.6 material frames Sol as the high-capability tier and reports results alongside GPT-5.5 and competing models. Those are vendor-published evaluation results, not an independent benchmark verdict, but they establish the product context for AWS's positioning. Bedrock is not presenting a separately trained AWS version of Sol; it is offering an AWS-operated access path with its own regional placement, identity controls, billing relationship, and service constraints.

For an existing OpenAI SDK application, AWS documents a relatively narrow migration: use the bedrock-mantle base URL, select the Amazon Bedrock model ID, and authenticate with a Bedrock API key or AWS credentials. The endpoint supports the Responses API and the same named reasoning-effort settings across the three tiers, which can reduce application-level branching when moving a workload between Sol, Terra, and Luna. AWS also documents prompt caching, including automatic caching and explicit breakpoints, but teams should measure cached-token fields in responses instead of assuming a cache hit from a stable prompt alone.

The governance details deserve as much attention as the compatibility story. AWS says calls run under IAM policies, can be kept in a selected region, and are logged in CloudTrail; it also says classifier-flagged traffic may be retained for up to 30 days for automated abuse detection, with retention controls available. That makes the offering potentially useful for organizations that need AWS account boundaries or regional processing, but it does not remove the need to assess the applicable OpenAI terms, the supported-region limit, quota planning, and whether an application can safely tolerate the latency and cost of higher reasoning effort. The available documentation supports the platform facts, not a universal claim that this route is preferable for every deployment.