Summary: Since September 2, 2026, Amazon Bedrock serves OpenAI GPT-5.6 models—Sol, Terra, and Luna—through global cross-Region inference, reachable from both Asia Pacific (Sydney) and Asia Pacific (Melbourne). For those building products, the practical question is how to actually invoke these models, and which details deserve attention. Two operational details stand out for production use.
If your team sits in Australia and has been waiting for a way to use OpenAI’s latest models without owning cross-Region routing, that option now exists. Since September 2, 2026, Amazon Bedrock serves OpenAI GPT-5.6 models—Sol, Terra, and Luna—through global cross-Region inference, reachable from both Asia Pacific (Sydney) and Asia Pacific (Melbourne). Your call goes to the Bedrock Runtime endpoint in your local Region, and Bedrock quietly picks a supported commercial AWS Region to handle it. What you gain is a far larger capacity pool, with no failover logic of your own to build.
For those building products, the practical question is how to actually invoke these models, and which details deserve attention. This post covers the setup, the three API paths, and a handful of operational tips such as prompt caching and authentication.
Three Models Behind a Single Endpoint
GPT-5.6 comes in three variants, and each one has a global inference profile ID that serves as the model identifier:
- Sol (
global.openai.gpt-5.6-sol): Aimed at workloads that lean hard on reasoning, coding, and agentic behavior. - Terra (
global.openai.gpt-5.6-terra): Keeps performance and cost in balance for everyday production work. - Luna (
global.openai.gpt-5.6-luna): Quick, inexpensive inference for apps with high volume and tight latency budgets.
All three take text and image inputs, generate text, and handle context windows of up to 1 million tokens. Either ap-southeast-2 (Sydney) or ap-southeast-4 (Melbourne) can reach them through the Bedrock Runtime endpoint. AWS chooses the destination Region on your behalf, so tracking which Region serves a given request is not something you need to do.
Before you begin, consult the Cross-Region inference support page, because profile membership and model availability are both subject to change. It is also worth confirming that the source Region is enabled on your AWS account, that your service control policy permits the GPT-5.6 profiles, and that your IAM permissions are sufficient.
Three Ways to Call the Models
The Bedrock Runtime endpoint exposes GPT-5.6 through three APIs. The OpenAI-compatible ones live under /openai/v1, and you can authenticate with either AWS SigV4 or a short-term Bedrock model inference API key. The AWS blog post demonstrates generating that key on the fly with the AWS Bedrock Token Generator for Python, which removes any need to store a static key.
OpenAI Responses API – Teams already on the OpenAI SDK only have to point the client at the Regional endpoint:
from aws_bedrock_token_generator import provide_token
from openai import OpenAI
region = "ap-southeast-2" # or "ap-southeast-4" for Melbourne
model_id = "global.openai.gpt-5.6-terra"
client = OpenAI(
base_url=f"https://bedrock-runtime.{region}.amazonaws.com/openai/v1",
api_key=provide_token(region=region),
)
response = client.responses.create(
model=model_id,
input="Explain Availability Zones in three bullet points.",
max_output_tokens=300,
)
print(response.output_text)For streaming, pass stream=True and loop over the events.
Chat Completions API – Apps written in the older Chat Completions style work as well. The blog’s example relies on the max_completion_tokens and reasoning_effort parameters, both familiar to OpenAI users.
Amazon Bedrock Converse API – When your codebase is built on AWS SDKs, reach for boto3 and the Converse API. Credentials resolve through the standard AWS chain, so nothing extra has to be configured. Streaming comes via converse_stream.
All three examples run against either Sydney or Melbourne—you swap the Region code and nothing else.
Prompt Caching and Codex Setup
Two operational details stand out for production use.
Prompt caching comes in two modes. Implicit caching is enabled by default and needs no code change. Explicit caching lets you declare reusable prefixes and cache keys, which can bring down costs on repetitive workloads. For examples, the blog points to the GPT-5.6 introduction post.
Codex can use these global inference profiles too. The blog’s validated setup pairs codex-cli 0.149.1 with GPT-5.6 Sol in Sydney. Organizations running Okta, Auth0, Microsoft Entra ID, Amazon Cognito, or IAM Identity Center have a sample OIDC auth helper that trades an OIDC token for temporary AWS credentials. You configure it with credential_process in ~/.aws/config, then point Codex at that profile in ~/.codex/config.toml. No API keys get stored, since requests are signed with SigV4.
Quota Management and Final Thoughts
Quota management is one subject the blog leaves open—it shows up as a section header with nothing filled in beneath it. That gap is something you will want to investigate yourself, particularly if high-volume usage is on your roadmap.
Viewed from Australia, this is a meaningful step: OpenAI models become usable through Bedrock while your API calls stay in your local Region. The tradeoff is that routing decisions belong to AWS, so latency and data residency may vary with wherever a request actually lands. If that matters for your workload, test against your own traffic patterns.
A sensible first move is to verify the inference profiles are active in your Region using the AWS CLI or console, then try the Responses API with a simple prompt. From there you can layer in caching and authentication as needed.