Bedrock cross-region inference just got a lot more interesting if you’re running OpenAI models. As of August 17, 2026, Amazon Bedrock supports GPT-5.6 Sol, Terra, and Luna directly on the bedrock-runtime endpoint, with the Responses, Converse, and Chat Completions APIs all available, plus two flavors of cross-Region routing: Global and Geo. If you’ve been treating OpenAI-on-Bedrock as a novelty, it’s time to stop. This is a real production routing layer now, and it changes how you should be architecting for throughput.
What Bedrock cross-region inference actually does
Cross-Region inference isn’t a new concept on Bedrock — it’s existed for other model families for a while — but extending it to OpenAI’s GPT-5.6 line closes a real gap. The core idea: instead of pinning your requests to capacity in one Region, you invoke a cross-Region inference profile, and Bedrock’s routing layer decides which underlying Region actually serves the request. You don’t provision anything Region by Region. You don’t watch CloudWatch dashboards praying us-east-1 has headroom during a demand spike. Bedrock handles it.
There are two routing modes, and they are not interchangeable:
- Global cross-Region inference — routes your request to any commercial AWS Region where the model is available. This gets you the broadest access to Bedrock capacity and the highest throughput ceiling during demand spikes. It is also priced lower per token than in-Region or Geo inference for OpenAI models, which is genuinely unusual — normally more flexibility costs more, not less.
- Geo cross-Region inference — constrains routing to a predefined geography. This release adds US Geo (US CRIS) support, meaning your request stays within US Regions even as it moves around for capacity. You get most of the throughput benefit with a residency guarantee.
Under the hood, both modes work through inference profiles — you don’t call a model ID directly, you call a profile ARN, and Bedrock’s control plane resolves it to whichever Region has capacity at that moment. This is the same mechanical pattern Bedrock uses for its other cross-Region-enabled models, just now wired up for Sol, Terra, and Luna.
The API surface: Responses, Converse, and Chat Completions
The other half of this launch is API breadth. GPT-5.6 models on bedrock-runtime now support three calling conventions:
- Responses API — OpenAI’s newer, more structured interface, useful if you’re porting code that already targets OpenAI’s native Responses format.
- Converse API — Bedrock’s own model-agnostic interface. If you’re calling Claude, Titan, or Llama on Bedrock today through Converse, you can slot GPT-5.6 into the same code path with minimal branching logic.
- Chat Completions API — the familiar OpenAI-shaped chat interface, for teams migrating existing OpenAI SDK code with the least friction.
Practically, this means you pick the API that matches your migration path, not the one Bedrock happens to support for a given model. That’s a meaningful shift from the early days of Bedrock model support, where API coverage was inconsistent across providers.
Opting in: picking an inference profile
To use cross-Region inference, you invoke the inference profile ARN instead of the base model ID. Here’s a boto3 example calling GPT-5.6 Terra through the Converse API using a Global cross-Region inference profile:
import boto3
client = boto3.client("bedrock-runtime", region_name="us-east-1")
response = client.converse(
modelId="arn:aws:bedrock:us-east-1:123456789012:inference-profile/global.openai.gpt-5-6-terra-v1:0",
messages=[
{
"role": "user",
"content": [{"text": "Summarize this incident report in three bullet points."}]
}
],
inferenceConfig={
"maxTokens": 1024,
"temperature": 0.3
}
)
print(response["output"]["message"]["content"][0]["text"])
For Geo (US CRIS) routing, you swap the profile prefix:
aws bedrock-runtime converse \
--region us-east-1 \
--model-id "arn:aws:bedrock:us-east-1:123456789012:inference-profile/us.openai.gpt-5-6-sol-v1:0" \
--messages '[{"role":"user","content":[{"text":"Draft a rollback plan for this deploy."}]}]'
Note the profile naming convention — global. for Global cross-Region inference, a geography code like us. for Geo. This mirrors the pattern Bedrock already uses for cross-Region inference profiles on other model families, so if you’ve wired this up before, the mental model transfers directly. Full syntax and Region availability lives in the Bedrock inference profiles documentation.
Model cards: Sol, Terra, and Luna
The three GPT-5.6 variants available on Bedrock are Sol, Terra, and Luna. Which one you pick is a workload decision, not a routing decision — cross-Region inference applies the same way regardless of variant, so pick the model tier based on latency and capability requirements for your use case, then layer Global or Geo routing on top. Don’t assume Global routing changes model behavior; it only changes where the request physically executes, not what the model returns.
Before any of this works, you still need standard Bedrock model access enabled for whichever GPT-5.6 variant you’re targeting — cross-Region inference profiles don’t bypass the model access gate, they sit on top of it. If you haven’t requested access to Sol, Terra, or Luna in the console yet, do that first; an inference profile pointing at a model you don’t have access to will just fail with an access-denied error, and the error message won’t tell you it’s an access problem versus a routing problem. Check your enabled models under the Bedrock model access page before debugging anything else.
Region availability is also worth confirming per variant rather than assuming parity. Not every GPT-5.6 variant is guaranteed to be available in every commercial Region on day one, even though Global cross-Region inference can route to any Region where the model is available. Run the list-inference-profiles call shown below against each Region you operate in before you commit application code to a specific profile ARN — the ARN itself is Region-scoped even when the routing behavior it enables is not.
Monitoring: CloudWatch and invocation logging
Because all of this runs through the standard bedrock-runtime endpoint, none of your observability tooling needs to change. Bedrock model invocation logging still captures full request and response payloads, deliverable to S3 or CloudWatch Logs, itemized by the model or inference profile you called. CloudWatch metrics — invocation counts, token counts, latency, throttles, errors — are emitted per inference profile, so you can tell at a glance whether your Global-routed traffic is hitting a different latency profile than your old in-Region calls. This matters more than it sounds: cross-Region routing means a single logical endpoint can now execute in a Region you didn’t explicitly choose, and if you’re not watching per-profile latency, you won’t notice a slow Region until customers do.
One CLI check worth adding to your runbook — list the inference profiles available to your account before you start swapping model IDs in production:
aws bedrock list-inference-profiles \
--region us-east-1 \
--query "inferenceProfileSummaries[?contains(inferenceProfileId, 'gpt-5-6')]"
That single call tells you whether Global, Geo, or both are actually live for the GPT-5.6 variant you want in your Region, before you find out the hard way in a deploy.
At Enkompass (enkompass.net), this is our home turf. We’ve reworked production Bedrock architectures to cut per-token spend without touching a single line of application logic, just by getting the routing layer right. If you’d rather have seasoned cloud people handle it instead of learning the hard way at 2 AM, let’s talk. Reach out through our contact page or call us directly at +1-828-436-5667 — we’ll help you turn this from a someday-project into a done-deal.