Bedrock adds cross-Region routing for GPT-5.6, pooling capacity across 25 regions
Requests for Sol, Terra and Luna can now draw on compute outside the region they were sent to, which AWS says holds throughput steady when a single region is saturated.
Original cover art, generated for this story. THE VISSION does not republish third-party press imagery.
- Amazon Bedrock now supports cross-Region inference for all three GPT-5.6 variants — Sol, Terra and Luna — routing requests beyond the region they were submitted to.
- AWS says the point is throughput under load: a request can draw on a broader compute pool rather than being bound by one region's spare capacity.
- US geographic profiles keep routing inside North America; global profiles span the US, Canada, eight European regions, twelve in Asia Pacific, two in the Middle East and São Paulo.
- All three variants take text and image input, carry a 1 million token context window, and support reasoning mode, server-side tool calling and prompt caching.
Amazon Bedrock has enabled cross-Region inference for OpenAI's GPT-5.6 family, letting a request submitted in one AWS region be served by capacity in another. AWS frames the change as a throughput measure rather than a latency one: by allowing requests "to draw on a broader pool of compute rather than being bound to one Region's available capacity," the service aims to hold performance steady when demand in a single region outruns what is provisioned there.
The routing is governed by geographic profiles, which is the part with compliance consequences. A US profile keeps both source and destination inside North America, drawing on US East in Northern Virginia and Ohio, US West in Oregon and Northern California, and Canada Central and Calgary. A global profile is far wider, spanning the United States, Canada, eight European regions, twelve across Asia Pacific, two in the Middle East and São Paulo — more than 25 regions in total.
All three GPT-5.6 variants are covered: Sol as the cost-efficient option, Terra balancing capability against price, and Luna as the most capable. Each accepts text and image input, carries a one-million-token context window, and supports reasoning mode, server-side tool calling and prompt caching. AWS pointed to its standard pricing page rather than announcing any change in rates alongside the routing feature.
Choosing between the two profile types is a data-residency decision dressed as a performance setting, and it is the kind of default that gets picked once during a prototype and inherited by production. A team with contractual obligations about where inference happens needs the US profile specifically; a team that selects the global profile for throughput has agreed, in effect, that a prompt submitted in Ohio may be processed in Singapore or São Paulo.