AWS has made Anthropic’s Claude Opus 5, Claude Sonnet 5 and Claude Haiku 4.5 available through Amazon Bedrock’s India geographic inference profile. Customers can use the models while processing data in AWS Regions within India, rather than relying only on the previously supported global cross-Region inference option.
The profile uses geographic cross-Region inference: an API request begins in the customer’s chosen source Region and Bedrock automatically sends it to available capacity in a permitted destination Region. For the India profile, routing is limited to ap-south-1 in Mumbai and ap-south-2 in Hyderabad. Prompts and generated results may move between those two Regions but, according to AWS, inference remains within India.
AWS says traffic between the Regions travels over its secure network with end-to-end encryption in transit. Customer data is not stored in the destination Region and remains in the source Region. Bedrock also applies zero data retention by default, meaning it does not store model inputs or outputs. AWS notes an exception: certain models may require human review when automated safety classifiers flag content.
Billing and quota usage are charged against the account in the source Region, regardless of which backend Region processes a request. Amazon CloudWatch and AWS CloudTrail likewise create their log entries only in the source Region, keeping monitoring and audit records associated with the point from which the API call originated.
Developers can access the models through the bedrock-runtime endpoint using Anthropic’s Messages API or Bedrock’s native InvokeModel and Converse APIs. Supported Bedrock capabilities include Guardrails and intelligent prompt routing. Programmatic calls require an India inference profile identifier, such as in.anthropic.claude-sonnet-5 or in.anthropic.claude-opus-5; the models can also be tried without code in the Bedrock console’s text playground.
Practical context: The practical distinction is that this is a geographically constrained pool spanning two AWS Regions, not a workload tied to capacity in a single Region. That broader pool may help maintain throughput during traffic peaks while preserving an India-only inference boundary. Organizations should still assess the source-Region storage model and the stated human-review exception against their own data-handling requirements.
Sources
Event date: 2026-09-30. Primary source date: 2026-09-30.
