AWS announced yesterday that Bedrock now runs OpenAI’s GPT-5.6 models, Sol, Terra, and Luna, directly on the bedrock-runtime endpoint, speaking the Responses API, Chat Completions API, and Converse API natively. Alongside that, it shipped cross-Region inference for those models: Geo routing that keeps requests inside a defined geography, and Global routing that spreads them across every commercial Region where the model is available.
This matters more than a typical “new model available” note. Since OpenAI models landed on Bedrock, they sat slightly apart from the rest of the platform: teams either translated their existing OpenAI SDK calls into Bedrock’s Converse format, or ran a parallel path outside Bedrock’s governance just to keep using the Responses API shape their code already expected. That workaround is now gone, and the routing decision it exposes, cost versus latency versus data residency, is one every team running GPT-5.6 at scale now has to make explicitly.
What actually changed
No more translation layer. If your backend calls OpenAI’s Responses API or Chat Completions API today, that exact request shape now runs against bedrock-runtime unchanged. You are not rewriting request bodies into Bedrock’s Converse schema just to get GPT-5.6 behind AWS’s account controls. Converse API support is still there too, so teams standardized on that abstraction across multiple model providers keep it.
Governance now applies uniformly. Because these calls run through bedrock-runtime, they inherit the same observability and cost surface as every other Bedrock model: invocation logging to S3 or CloudWatch Logs, CloudWatch metrics for latency, throttles, and token counts, and itemized spend in Cost Explorer and the Cost and Usage Report. For any org that mandated Bedrock for AI spend visibility, GPT-5.6 usage no longer needs a separate reporting path bolted on afterward.
Routing is now a real decision, not a default. Geo cross-Region inference keeps requests inside one geography, with a new US Geo option (US CRIS) shipping in this launch, so data stays within that boundary while you still get multi-Region capacity. Global cross-Region inference routes to any commercial Region where the model runs, which gives the highest available throughput during demand spikes and is priced lower per token than in-Region or Geo routing. In-Region inference is still the default if you do neither.
Picking a routing mode
The tradeoff is genuinely three-way: cost, throughput headroom, and data residency pull in different directions. Answer the two questions below and see which mode fits.
About the demo
This is a decision aid based on AWS’s published behavior for each routing mode, not a pricing calculator. AWS did not publish exact per-token deltas between Geo and Global at launch.Answer both questions and click Recommend.
Where this leaves you
If you already run GPT-5.6 on Bedrock through the Converse API, nothing breaks, you just gained two new routing options worth revisiting. If you were maintaining a request-translation shim to keep using OpenAI’s native API shapes, that shim is now dead weight, delete it and point the same request bodies at bedrock-runtime. And if you were running OpenAI calls outside Bedrock entirely just to avoid the translation cost, this removes the last real argument for that split: you now get native request shapes and unified cost attribution in the same place.
My take: this is AWS treating OpenAI models as first-class Bedrock citizens rather than a bolted-on integration, and the routing split is the more interesting piece long-term. Geo cross-Region inference is a quiet acknowledgment that “just use Global for everything” was never going to satisfy regulated industries, and giving that a named, documented mode (rather than leaving teams to pin Regions by hand) is the right call. Worth checking whether your GPT-5.6 traffic is still hard-pinned to a single Region out of habit rather than requirement, that is capacity you are leaving on the table during the next demand spike.