OutYet reporting
AWS documents a cross-Region path for GPT-5.6 Sol, but its model card has not caught up
AWS describes new geographic and global Bedrock inference profiles for GPT-5.6 Sol, while the current Sol model card still lists only in-Region access. That gap matters for teams planning capacity, residency controls, and IAM policies.
AWS published guidance on August 20 for running GPT-5.6 Sol, Terra, and Luna through Amazon Bedrock cross-Region inference profiles. The guidance describes more than 25 AWS Regions and says the three general-purpose variants can be called through OpenAI Responses and Chat Completions formats or Bedrock Converse. For Sol users, the change is operational rather than a claim about new model behavior: the relevant decision is whether a workload should draw on capacity beyond the source Region.
The implementation uses an inference-profile identifier instead of a raw model ID. AWS gives geographic IDs such as us.openai.gpt-5.6-sol for routing inside a defined geography, and global IDs such as global.openai.gpt-5.6-sol for routing across eligible commercial Regions. Existing OpenAI SDK integrations can point at Bedrock's OpenAI-compatible endpoint and substitute the profile ID, while applications already standardized on Bedrock can use Converse. That makes this primarily an endpoint, authentication, policy, and observability migration question.
There is an important documentation inconsistency to resolve before treating those identifiers as a production contract. AWS's detailed August 20 post names both geographic and global Sol profile patterns, but the AWS Sol model card currently shows the raw openai.gpt-5.6-sol ID and marks Geo and Global inference IDs as unsupported. The sources do not explain the mismatch. It may reflect a lagging model card, a different service surface, or account and Region conditions that the card does not express. The safest reading is that teams should verify the profile in their intended Bedrock Region rather than infer availability from either page alone.
AWS's own caveats make the profile choice consequential. Geographic routing is presented for workloads that must keep processing within a specified geography; global routing has the broadest capacity pool but may process data in any eligible commercial Region. AWS also says quota pools are separate by profile, output tokens count at a 10x burndown rate for GPT-5.6 quota calculations, and latency can vary with the destination Region. Teams should therefore test with their actual prompt and output mix, review IAM and service-control policies, and monitor by inference-profile ID. The sources document the routing mechanics, not a universal latency or cost outcome for a particular application.