On August 27, 2026, Amazon Bedrock made OpenAI GPT-5.6 Terra and Luna available for inference processed within India. Workloads with local processing requirements can now use both models through Bedrock's operating environment.

“In-country” does not mean “always in Mumbai,” however. The feature is geographic cross-Region inference that draws on capacity in both Mumbai and Hyderabad.

That distinction drives the adoption decision. Once the country boundary and two-Region routing are treated as one design, the required API, permission, storage, and audit choices become much clearer.

India Geo profiles choose between Mumbai and Hyderabad capacity

With India geographic cross-Region inference, callers use either Asia Pacific (Mumbai), ap-south-1, or Asia Pacific (Hyderabad), ap-south-2, as the source Region. Amazon Bedrock chooses an inference destination within those two Regions according to capacity.

An inference profile bundles the model with the Regions to which a request may be routed, replacing client-side capacity distribution. The profile IDs are in.openai.gpt-5.6-terra for Terra and in.openai.gpt-5.6-luna for Luna.

This builds on an existing Bedrock mechanism. AWS launched cross-Region inference on August 27, 2024 to draw on compute across Regions during traffic bursts. The India Geo profiles apply that capacity mechanism inside an India-only geographic boundary.

The Amazon Bedrock inference options differ primarily in where processing may occur.

Use an in. profile when inference must remain in India. India Geo is not a substitute for strict single-Region processing: prompts and outputs remain in the country, but they can move between Mumbai and Hyderabad.

On bedrock-runtime, the endpoint and modelId define the boundary

AWS recommends the bedrock-runtime endpoint for new applications. It supports the OpenAI-compatible Responses and Chat Completions APIs as well as the Amazon Bedrock Converse API, so teams can choose an entry point that fits their existing client stack.

For the OpenAI SDK and Responses API, point the client at an India Region and pass an India Geo profile ID as model. This minimal example calls Terra from Mumbai.

from aws_bedrock_token_generator import provide_token
from openai import OpenAI

REGION = "ap-south-1"
MODEL_ID = "in.openai.gpt-5.6-terra"

# Derive a short-lived token from AWS credentials instead of storing a static API key.
client = OpenAI(
    base_url=f"https://bedrock-runtime.{REGION}.amazonaws.com/openai/v1",
    api_key=provide_token(region=REGION),
)

# The India Geo profile limits inference routing to the in-country boundary.
response = client.responses.create(
    model=MODEL_ID,
    input="Extract the payment due date and total amount from the invoice and return JSON.",
    max_output_tokens=512,
)

print(response.output_text)

Existing bedrock-mantle workloads remain supported, but adopting the India Geo profiles requires pointing them at bedrock-runtime in Mumbai or Hyderabad. Teams standardized on AWS SDKs and SigV4 can instead pass the same profile ID to the Converse API.

Terra and Luna serve different roles inside the same boundary

Both models have a one-million-token context window, accept text and image input, and produce text output. AWS positions Terra as a balanced model for general production work and Luna as a faster, lower-cost option for high-volume classification, summarization, routing, and real-time workloads.

Those descriptions are selection starting points, not workload-specific guarantees. Evaluate both models with representative inputs and the same quality criteria, then compare output quality, latency, and token use. Supporting a long context window also does not mean that sending the maximum context on every request is optimal.

IAM and SCP policies must permit both India Regions

The application-level change is mainly an endpoint and model ID, but organization-level Region controls must cover both destinations. The IAM requirements for geographic cross-Region inference require access to the inference profile, the foundation model in the source Region, and the foundation model in every destination Region.

For India Geo, both ap-south-1 and ap-south-2 must be permitted. If an AWS Organizations SCP blocks either destination, cross-Region inference fails even when the source Region itself is allowed. The bedrock:InferenceProfileArn condition can limit which profiles are usable while the two Regions remain available.

In-country processing still needs separate retention and audit decisions

Data residency defines where inference is processed; it does not unconditionally mean that nothing is stored. Amazon Bedrock uses zero data retention by default, but for GPT-5.6 Terra and Luna, traffic flagged by abuse-detection classifiers can be retained for up to 30 days for offline abuse detection. With cross-Region inference, retained inputs and outputs are stored in the destination Region that processed the request.

Explicit storage paths also appear when an application uses store=True for server-side Responses state or enables customer-controlled invocation logging. Model invocation logging is disabled by default; when enabled, it can deliver request data, response data, and metadata to CloudWatch Logs or Amazon S3. Retention, encryption, and access controls for those logs must be designed separately from the inference boundary.

For audit evidence, the additionalEventData.inferenceRegion field in CloudTrail identifies the destination that handled a request. With an India profile it is ap-south-1 or ap-south-2, providing an operational record that processing stayed in the country.

Adopt India Geo as an in-country, two-Region design

India Geo fits workloads that cannot send inference outside India but may be processed in either Mumbai or Hyderabad. A system that requires one fixed Region cannot treat this launch alone as satisfying that requirement.

A practical adoption sequence is:

1. State whether the compliance boundary is “within India” or “one AWS Region.”
2. Test representative data on bedrock-runtime with in.openai.gpt-5.6-*, comparing Terra and Luna.
3. Permit both Mumbai and Hyderabad in IAM and SCP policies, then scope access to the intended profile.
4. Incorporate ZDR exceptions, store, invocation logs, CloudTrail evidence, and current pricing into the production design.

The practical value of this launch is larger than model availability alone. It lets an organization treat two Indian Regions as one inference-capacity pool and enforce the same country boundary through model IDs, permissions, and audit records. For workloads designed around that boundary, it offers a credible way to combine data residency with scale.

Source

• Title: Introducing OpenAI models on Amazon Bedrock for in-country inferencing in India
• URL: https://aws.amazon.com/blogs/machine-learning/introducing-openai-models-on-amazon-bedrock-for-in-country-inferencing-in-india/