AWS adds cross-Region inference for GPT-5.6 models on Bedrock

Amazon Web Services says Amazon Bedrock now offers OpenAI GPT-5.6 Sol, Terra, and Luna through cross-Region inference profiles in more than 25 AWS Regions. Its announcement says customers can use US geographic or global profiles, call the models through either the OpenAI or Converse API, and configure the associated IAM, quota, and monitoring settings. This is an availability and routing change in Bedrock, not evidence of a new GPT-5.6 model release or a change to its underlying capabilities.

The relevant mechanism is an Amazon Bedrock inference profile, which binds a foundation model to a set of destination Regions. A geographic profile keeps processing within the selected geography, while a global profile may route a request to any supported commercial AWS Region. AWS documents the distinction as a data-residency and cost decision: geographic routing is the fit for location constraints, while global routing trades geographic restriction for an advertised approximate 10% saving. AWS also states that cross-Region traffic remains on its network, is encrypted in transit, and can be traced in CloudTrail through the processed inference Region.

For GPT-5.6 Sol users, the operational change is separate from the model's direct OpenAI API surface. OpenAI describes Sol as the frontier tier in the GPT-5.6 family, with a 1,050,000-token context window, 128,000 maximum output tokens, configurable reasoning effort, and support for both Responses and Chat Completions. Bedrock's new profiles therefore do not alter those documented model limits; they give AWS-based deployments a different routing and governance layer around requests. Teams moving an existing direct-API workload should validate API compatibility, quotas, logging, and billing rather than assume the new route is a transparent endpoint substitution.

The practical benefit is potential resilience and broader regional access for applications that already standardize on Bedrock, especially where a single Region's capacity is a bottleneck. The limitations are material: global profiles can process requests outside a chosen geography, inference profiles do not support Provisioned Throughput, and service-control policies must permit the profile's destination Regions or the special unspecified Region condition for global routing. AWS says CloudTrail records the source-Region request and exposes the inference Region, so production teams should use that field to verify where workloads actually ran. The announcement does not provide workload-specific latency, quality, or capacity measurements, so those remain deployment questions rather than established performance gains.