India cross‑Region inference for OpenAI models on Bedrock: practical impact for engineers

Amazon Bedrock now exposes the OpenAI GPT‑5.6 Terra and Luna models through India‑specific cross‑Region inference profiles, ensuring that all request payloads and responses stay within the Indian geography. This change lets teams that must keep data inside India run large‑context, multimodal workloads without manually provisioning capacity in each region.

What changed in Bedrock

Two new inference profiles – in.openai.gpt-5.6-terra and in.openai.gpt-5.6-luna – are available in the Asia Pacific (Mumbai) ap-south-1 and Asia Pacific (Hyderabad) ap-south-2 Regions. Both models support a 1‑million‑token context window, accept text and image inputs, and return text. The Bedrock runtime endpoint is the recommended entry point, exposing the native InvokeModel and Converse APIs as well as OpenAI‑compatible Responses and Chat Completions APIs.

Implications for architecture and operations

Cross‑Region inference works by routing a request from the source Region to any of the Indian Regions that have spare capacity. The routing decision is transparent to the caller; billing and quota are recorded against the source Region only. This model simplifies capacity planning because you no longer need to pre‑size each Region individually.

• Endpoint selection: Use the bedrock-runtime endpoint in either ap-south-1 or ap-south-2. The profile ID you pass determines which model and which set of Regions the request may travel through.
• Monitoring: CloudWatch metrics and CloudTrail logs appear only in the source Region, keeping observability consolidated.
• Fail‑over: If one Indian Region experiences a capacity shortage, Bedrock automatically fails over to the other Region without client‑side changes.

Security and data residency considerations

The India‑specific profiles guarantee that data never leaves the country; payloads may move between Mumbai and Hyderabad but remain on the Amazon network, encrypted in transit. Bedrock follows a zero‑data‑retention model, meaning inputs and outputs are not persisted by default. An exception exists for content flagged by the built‑in abuse‑detection classifiers, which is retained for offline analysis.

Because logs are generated only in the source Region, audit trails remain within the Indian geography, simplifying compliance with local regulations. However, practitioners should be aware that any retained flagged content will be stored according to the service’s abuse‑detection policy.

Related CloudNinjas coverage: AWS.

What This Means For Practitioners

Engineers can now integrate GPT‑5.6 Terra or Luna into pipelines that process long documents, extensive codebases, or mixed text‑image workloads while satisfying Indian data‑residency mandates. The primary actions are:

1. Update your Bedrock client to call the bedrock-runtime endpoint in ap-south-1 or ap-south-2 and specify the appropriate in.openai.gpt-5.6-* profile.
2. Validate that your monitoring stack captures CloudWatch and CloudTrail events from the source Region only.
3. Review the abuse‑detection retention policy if your workload may generate flagged content.
4. Consider whether the global cross‑Region profiles (global.*) are needed for capacity beyond the Indian geography, keeping in mind the trade‑off of data leaving the country.

Going forward, keep an eye on Bedrock’s regional model availability matrix and any updates to the zero‑data‑retention handling of flagged content, as those will affect both compliance posture and operational monitoring.