GPT-5.6 Sol gains an AWS-native path, but the integration changes operations more than the model

AWS says its Bedrock Mantle service can invoke GPT-5.6 Sol, Terra, and Luna through the OpenAI Responses API. Its published model table places Sol in autonomous coding, security research, scientific analysis, and deep multi-step reasoning; Terra is the general-purpose balance of reasoning and cost; Luna targets high-volume, latency-sensitive work. This is a new deployment option for the existing GPT-5.6 family, not an AWS-specific model announcement. The published regional table lists Sol in US East (N. Virginia) and US East (Ohio), while Terra and Luna also list US West (Oregon).

The integration is OpenAI-compatible, but it is not operationally identical to calling OpenAI directly. AWS instructs teams moving an existing OpenAI SDK application to replace the base URL with a regional `bedrock-mantle` endpoint, use an AWS-specific model ID, and authenticate with a Bedrock API key or AWS credentials. AWS also says all three models accept text and image input, return text, expose a 272K-token context window, and support six reasoning levels from `none` through `max`. That preserves much of an existing Responses API integration while moving authentication and endpoint ownership into AWS.

OpenAI presents the same three-tier family as a capability and price ladder: Sol at $5 input and $30 output per million tokens, Terra at $2.50 and $15, and Luna at $1 and $6. AWS says its Bedrock pricing matches OpenAI's first-party rates and that usage counts toward existing AWS commitments. The useful comparison, then, is not a claim that Bedrock improves Sol's underlying reasoning. It is a procurement and deployment choice: a team can keep the model tier while placing inference under AWS account controls and a familiar cloud billing relationship.

AWS describes those controls as IAM-governed calls that can run in a chosen VPC, are logged in CloudTrail, and support in-region inference. Its documentation also sets an important limitation: classifier-flagged traffic may be retained for up to 30 days for automated abuse detection, and retained inputs and outputs are processed by AWS rather than shared with OpenAI unless a customer opts in. Teams with residency or retention requirements should therefore validate the applicable Bedrock retention mode and the supported region for Sol, rather than infer that an OpenAI-compatible endpoint alone satisfies their policy.

The most concrete cost feature is prompt caching. AWS says cached input receives a 90% discount, cache writes cost 1.25 times the uncached input rate, cached material is retained for at least 30 minutes, and an explicit cache checkpoint needs a prefix of at least 1,024 tokens. It also warns that a cache hit is not guaranteed on every identical-looking request and directs users to inspect `cached_tokens` in each response. That makes the feature relevant for agent loops with stable system prompts or tool definitions, but it is not a substitute for measuring real cache-hit rates and end-to-end latency.

What is verified here is the Bedrock interface, supported regions, configuration pattern, and AWS-stated data-handling and caching terms. OpenAI's public GPT-5.6 page separately describes Sol as the premium tier and publishes its own benchmark and pricing claims, but neither source establishes that every workload will benefit from moving to Bedrock. Technical users should run representative evaluations, especially if they rely on long-lived agent context, a non-US deployment region, or a zero-retention policy. The new path is most consequential where AWS governance and existing cloud commitments outweigh the added endpoint and credential migration work.