AWS adds India-bound Bedrock paths for GPT-5.6 Terra and Luna

On August 27, Amazon Web Services said Amazon Bedrock had made OpenAI GPT-5.6 Terra and GPT-5.6 Luna available through India geographic cross-Region inference. The profiles are named `in.openai.gpt-5.6-terra` and `in.openai.gpt-5.6-luna` and are presented for workloads that need processing inside India. AWS describes both variants as accepting text and image inputs, producing text output, and offering a one-million-token context window. The announcement concerns a regional serving and routing option for the two named variants rather than a newly named GPT-5.6 model.

AWS describes cross-Region inference as a capacity mechanism that can route a request across regions without customers managing capacity in each one. For the India geographic profiles, that choice is confined to the Mumbai and Hyderabad regions, although prompts and outputs may move between those two regions. AWS says Bedrock does not store inputs or outputs by default under its zero-data-retention model, but content flagged by its automated abuse-detection classifiers, including for GPT-5.6, can be retained for offline abuse detection. That caveat is material for teams interpreting the feature as an absolute no-retention guarantee.

The India-specific profiles differ from AWS global cross-Region inference profiles, which can route an India-originated request to supported commercial AWS regions worldwide and support GPT-5.6 Sol as well as Terra and Luna. The country-bounded `in.` profiles are documented only for Terra and Luna, so they are not a locality-preserving option for every member of the family. Teams with a requirement to keep processing within India must choose the geographic profile rather than a global profile, accepting a smaller regional capacity pool in exchange for the routing boundary.

Implementation is designed to fit software already using OpenAI-style calls: AWS documents Responses and Chat Completions compatibility through Bedrock endpoints, with the inference-profile identifier passed as the model value. It also supports the Bedrock Converse API for applications using AWS SDK authentication. The practical limits are operational rather than cosmetic: production access requires appropriate Bedrock permissions for the profile and foundation models in both India regions, and an organization-level region policy that blocks either destination can make requests fail. AWS also documents prompt caching for these profiles, with a stated 90 percent discount on cached reads, but that benefit depends on repeated long prefixes and does not remove the need to validate application behavior, availability, and account policy in the target environment.