GPT-5.6 Sol gets a Bedrock capacity route, while Ultrafast remains a limited preview

AWS says GPT-5.6 Sol, Terra, and Luna can now be invoked through Amazon Bedrock cross-Region inference profiles in more than 25 AWS Regions. The profiles let an application call one model identifier while Bedrock routes the request to eligible capacity in another Region. For Sol users already running on Bedrock, this is a deployment change: it increases the set of capacity options without changing the underlying model family or making any claim about a new model release.

The timing matters because OpenAI had announced a different Sol service only a week earlier. On August 13, OpenAI described Ultrafast as a limited-preview API tier for a select group of customers, powered by Cerebras, with a stated ceiling of 750 output tokens per second and up to 14 times Standard processing speed. AWS's August 20 announcement instead concerns the ordinary Bedrock access path for all three GPT-5.6 general-purpose variants, with its own API and routing controls.

The two changes solve different constraints. OpenAI's preview is framed around rapid generation for interactive workloads, whereas Bedrock's cross-Region inference is framed as a capacity mechanism that can improve throughput by drawing on compute beyond a single Region. AWS offers a US geographic profile that keeps processing within its predefined geography and a global profile that can route among supported commercial Regions. A team cannot treat those modes as interchangeable: the global profile may process data in another eligible Region, while the geographic profile deliberately narrows that routing boundary.

For technical users, the practical question is therefore not simply whether Sol is faster. Teams using Bedrock can retain an OpenAI-compatible Responses or Chat Completions integration, or use Bedrock Converse, while selecting an inference-profile identifier instead of a raw model ID. AWS says permissions must cover the profile and the foundation model in every possible destination Region, and that the processed Region is recorded in CloudTrail. OpenAI also says Ultrafast access is still limited and will expand only as capacity grows, so neither announcement establishes general access to that preview tier or guarantees a particular end-to-end workload latency.