AWS Adds Cross-Region Routing for OpenAI Models on Bedrock

Amazon Bedrock customers running OpenAI's GPT-5.6 models can now let AWS decide which region handles each inference request. The company announced on August 17, 2026 that Cross-Region inference is extending to the OpenAI model family, covering three variants named Sol, Terra, and Luna.

The feature works by automatically distributing requests across multiple AWS Regions rather than pinning traffic to wherever a customer happens to deploy. That removes a common bottleneck: teams no longer need to manually provision or shift capacity when demand spikes in a single region.

AWS is offering two routing modes. Global cross-region inference can serve a request from any commercial AWS Region where the model is available, and it comes with lower per-token pricing than either in-region or Geo-based inferencing. Geo cross-region inference keeps traffic within a defined geography instead, and the update adds a new US Geo option (US CRIS) for customers who need requests to stay inside US infrastructure for compliance or latency reasons.

The GPT-5.6 models are accessible through the bedrock-runtime endpoint and support three separate interfaces: the Responses API, the Converse API, and Chat Completions. That gives developers flexibility to integrate OpenAI models using whichever calling convention their existing codebase already expects, rather than rewriting integration logic around a new format.

What this changes operationally

Because the models run through bedrock-runtime, they inherit the same governance tooling Bedrock already applies to its other model providers. That includes account-level access controls, usage logging pushed to S3 and CloudWatch Logs, CloudWatch metrics for monitoring, and line-item cost tracking in AWS Cost Explorer.

For engineering teams, that means OpenAI model usage on Bedrock can be audited and billed with the same processes already used for Anthropic, Amazon's own Titan models, or any other Bedrock-hosted provider. There's no separate dashboard or logging pipeline to stand up.

Cross-Region inference for these OpenAI models rolls out to every AWS Region where the models themselves are already offered, so availability tracks the existing model footprint rather than launching in a narrower subset of regions first.

The practical upside is cost and reliability rather than new model capability: the announcement doesn't claim any change to GPT-5.6's underlying performance, only to how efficiently and cheaply Bedrock can route traffic to it at scale.