As organizations in highly regulated sectors like finance and healthcare face increasing pressure to balance generative AI adoption with strict data residency laws, the infrastructure supporting these models has become a critical point of friction. This week, Amazon Bedrock addressed this challenge by launching the GPT-5.6 model family—specifically the Terra and Luna variants—within the Asia Pacific (Mumbai) ap-south-1 and Asia Pacific (Hyderabad) ap-south-2 regions.
Data Sovereignty and Regional Routing
The core of this update is the introduction of India-specific inference profiles. By using the `in.` prefix in the model identifier, organizations can ensure that all inference requests are processed exclusively within the two specified Indian regions. This architecture prevents data from crossing national borders, satisfying local compliance requirements for data residency. When a request is initiated, Amazon Bedrock automatically routes the traffic between `ap-south-1` and `ap-south-2` based on real-time capacity, ensuring high availability without violating geographical constraints.
These models support a 1-million token context window, allowing for extensive document analysis within the secure boundary. While data is encrypted during transit, Amazon Bedrock maintains a zero-data-retention policy, meaning inputs and outputs are not stored by the service. Exceptions are limited to data flagged by the platform's automated abuse detection classifiers, which may be retained for offline analysis as outlined in the Amazon Bedrock User Guide.
Operational Efficiency via Cross-Region Inference
The system utilizes cross-region inference to maintain consistent throughput during traffic spikes. By pooling capacity between the Mumbai and Hyderabad regions, the service avoids the bottlenecks associated with single-region deployments. Crucially, this does not complicate administrative overhead; billing, quotas, and monitoring logs are consolidated at the source region. Developers can track performance and costs via Amazon CloudWatch and AWS CloudTrail without needing to aggregate logs from multiple disparate locations.
Choosing Between Global and Local Profiles
Amazon Bedrock now distinguishes between two primary deployment profiles. The `global.` profile is designed for maximum capacity, routing requests across the worldwide AWS infrastructure to optimize performance. In contrast, the `in.` profile is strictly limited to the Indian geographical boundary. For developers, the integration process remains seamless. Because the platform supports standard AWS credentials and Bearer tokens, existing applications using the OpenAI SDK can be migrated to these regional endpoints simply by updating the model ID and endpoint configuration.
client.chat.completions.create(
model="in.model-id-example",
messages=[{"role": "user", "content": "Hello!"}],
reasoning={"effort": "low"}
)Cost Optimization with Prompt Caching
To further improve efficiency, Amazon Bedrock has enabled prompt caching for these models. For requests containing a prefix of at least 1,024 tokens, users can realize up to 90% cost savings. This is particularly effective for RAG (Retrieval-Augmented Generation) workloads or agentic workflows where system instructions and knowledge base extracts are repeatedly sent. The caching mechanism, which supports both explicit and implicit modes, operates entirely within the Indian data residency boundary, ensuring that cost-saving measures do not compromise compliance.
import boto3client = boto3.client('bedrock-runtime', region_name='ap-south-1')
response = client.invoke_model(
modelId='arn:aws:bedrock:ap-south-1::foundation-model/gpt-5-6-terra',
body=json.dumps({
"prompt": "캐시될 시스템 명령어 및 지식 베이스",
"inferenceConfig": {
"cache": {
"mode": "explicit"
}
}
})
)
By decoupling the choice of data residency from the underlying model performance, Amazon Bedrock allows developers to maintain strict governance while leveraging the full capabilities of the GPT-5.6 architecture.
