AWS adds India-bound routing for GPT-5.6 Terra and Luna on Bedrock

Amazon Web Services says Amazon Bedrock now supports GPT-5.6 Terra and GPT-5.6 Luna through India geographic cross-Region inference. The service exposes the profiles as `in.openai.gpt-5.6-terra` and `in.openai.gpt-5.6-luna`, with requests routed between the Mumbai and Hyderabad AWS Regions. AWS says the arrangement is intended for customers that need inference processing to remain in India while still drawing on capacity across those Regions.

The operational change is more specific than selecting an India endpoint. AWS describes an inference profile as the model identifier that also defines the Regions eligible to serve a request. It says billing, quotas, CloudWatch, and CloudTrail records remain associated with the source Region even when capacity sends work to the other Indian Region. That gives an operator one regional place to inspect usage while preserving a record of the destination inference Region in CloudTrail.

The new profiles are a deployment and routing option for the existing Terra and Luna variants, not another member of the GPT-5.6 family. OpenAI describes Terra as the version intended to balance intelligence and cost, with a 1.05 million-token context window and configurable reasoning effort. AWS says the Bedrock profiles support the OpenAI Responses and Chat Completions formats as well as Bedrock's Converse API, so an application can choose an integration boundary without changing the model family it uses.

For a team already using the OpenAI SDK, AWS documents redirecting the client to the Bedrock Runtime endpoint and passing the India profile ID as the model parameter. That can make the change relevant to workloads that need AWS identity, logging, and regional controls alongside an OpenAI-compatible request format. The same source says the profiles also support prompt caching, which could matter for repeated long-context retrieval or agent prompts, subject to the workload's actual token pattern.

The residency boundary has limits that technical users should evaluate rather than assume away. AWS says prompts and outputs can move between Mumbai and Hyderabad, and both Regions must be allowed by any Service Control Policy or calls can fail. It also says that, although Bedrock does not store model inputs or outputs by default, content flagged by automated abuse-detection classifiers for GPT-5.6 can be retained for offline abuse detection. The announcement therefore supports a bounded India-processing claim, not a blanket promise that no content is ever retained or that every surrounding system is region-local.