AWS Adds India-Only Cross-Region Inference for OpenAI GPT-5.6 Terra and Luna

Amazon Bedrock adds an India-only routing boundary for GPT-5.6

AWS announced on August 27, 2026 that Amazon Bedrock now supports OpenAI GPT-5.6 Terra and Luna through India geographic cross-Region inference. The new profiles are designed for workloads that need more capacity than a single AWS Region can provide while still keeping inference processing inside India.

The two India profile identifiers are in.openai.gpt-5.6-terra and in.openai.gpt-5.6-luna. AWS says requests submitted from its Mumbai or Hyderabad Regions can be routed between Asia Pacific (Mumbai), ap-south-1, and Asia Pacific (Hyderabad), ap-south-2, depending on available capacity.

That is different from Bedrock's global inference profiles, which can send a request to supported commercial AWS Regions outside India. For teams with a local-processing requirement, the important control is therefore the India-prefixed inference profile, not merely calling Bedrock from an India endpoint.

What the GPT-5.6 India profiles support

AWS says both Terra and Luna support a 1-million-token context window, accept text and image input, and produce text output. The models can be called through OpenAI-compatible Responses and Chat Completions APIs as well as the Bedrock Converse API. Streaming is also available.

For new applications, AWS recommends the bedrock-runtime endpoint because it is where capabilities such as geographic cross-Region inference, Bedrock Guardrails and intelligent prompt routing are exposed. Existing Bedrock Mantle workloads remain supported, but the India geographic profiles are a Runtime feature.

The practical benefit is not a new model capability by itself. Cross-Region inference is primarily a capacity and routing mechanism: an application can draw on compute in two India Regions rather than being limited to the capacity of one Region. That can reduce the operational burden of manually balancing traffic between Mumbai and Hyderabad during peaks.

Data residency has an important implementation detail

AWS states that the India geographic profiles keep model inference processing within India and that traffic between Mumbai and Hyderabad is encrypted in transit across the Amazon network. CloudTrail can record the destination inference Region, providing an auditable signal of whether a request was handled in ap-south-1 or ap-south-2.

AWS also describes Amazon Bedrock as using a zero-data-retention security model by default. However, its announcement includes a qualification for GPT-5.6: content flagged by Bedrock's automated abuse-detection classifiers can be retained for offline abuse detection under the applicable Bedrock policy. Teams evaluating the feature for regulated workloads should therefore read the current Bedrock data-retention and abuse-detection documentation rather than interpreting 'within India' as meaning that no policy-based retention can ever occur.

Existing OpenAI SDK applications can use the Bedrock endpoint

One notable developer feature is API compatibility. AWS shows existing OpenAI SDK clients being pointed at the Bedrock OpenAI-compatible Runtime endpoint and supplied with an India inference profile as the model identifier.

Authentication can use standard AWS credentials or a Bedrock API key. For production use, AWS recommends short-lived credentials rather than embedding long-lived static secrets. The same India profiles also work with Bedrock prompt caching and its monitoring stack.

AWS says the models expose multiple reasoning-effort settings and can maintain server-side conversation state when Responses API calls are created with storage enabled. Those capabilities matter for agent and long-context applications, but they do not change the geographic routing rule: an in. profile is what restricts processing to the India geography.

Codex can also use the India Bedrock Runtime profile

The AWS post includes a Codex configuration example using the Amazon Bedrock Runtime provider with an India GPT-5.6 profile. Organizations can rely on normal AWS credential resolution or federated short-term credentials. This makes the announcement relevant beyond conventional application APIs: coding-agent deployments can also operate through the same India-restricted inference layer.

Why this matters

The release is a useful example of cloud providers separating model choice from data-processing geography. The same model family can be offered through global routing for maximum capacity or through a narrower geographic profile for organizations with residency constraints.

For Indian financial-services, healthcare, public-sector and enterprise teams, the key question is therefore not simply whether GPT-5.6 is available on Bedrock. It is whether the selected inference profile, IAM policy, logging configuration, retention policy and surrounding application architecture satisfy the organization's actual compliance requirements.

AWS's India profiles give those teams a more explicit infrastructure control: Terra and Luna workloads can scale across Mumbai and Hyderabad without intentionally routing inference outside the country.