Amazon Web Services has opened access to OpenAI’s GPT-6 Sol and GPT-6 Luna on Amazon Bedrock through global cross-Region inference from five Asia Pacificpore), the company said in an AWS Machine Learning Blog post published October 12, 2026
The post is a configuration walkthrough for teams in Thailand, Malaysia, Singapore, Indonesia, and Taiwan, covering how to place the call, set permissions, and monitor inference from theion inference, which it says can throttle when demand exceeds local capacity
Global Cross-Region Inference Profiles
Three profile IDs are listed on the bedrock-runtime endpoint. Callers reach GPT-6 Sol and GPT-6 Luna through global.openai.gpt-6-sol and global.openai.gpt-6-luna from Asia Pacific (Singapore) ap-southeast-1, Asia Pacific (Jakarta) ap-southeast-3, Asia Pacific (Malaysia) ap-southeast-5, Asia Pacific (Thailand) ap-southeast-7, and Asia Pacific (Taipei) ap-east-2. GPT-6 Astra is reachable through global.openai.gpt-6-astra from Singapore alone. Destination Regions are the supported commercial AWS Regions where each model is deployed, and the Amazon Bedrock console lists all three profiles under system-defined inference profiles.
Under global cross-Region inference, the caller passes a profile ID in place of a model ID. The the destination Region for processing, and billing, quota tracking, AWS CloudTrail, and Amazon CloudWatch remain in the to Amazon Bedrock capacity across Regions
AWS characterizes the three models by workload: GPT-6 Sol for reasoning and coding on complex tasks, GPT-6 Luna for focused, repeatable tasks at high volume, and GPT-6 Astra for complex reasoning, coding, and long-document work. All three accept text and image input, client-side tool use, and prompt caching, according to the post. They differ on reasoning effort: Sol and Luna accept settings from none through max, while Astra accepts the same levels except none.
Access Paths and Coding Clients
Each model supports three access paths on bedrock-runtime: the Amazon Bedrock Converse API, the OpenAI Responses API, and the OpenAI Chat Completions API. The OpenAI-compatible APIs are called on /openai/v1 paths rather than through the AWS SDKs and accept either Signature Version 4 signing or an Amazon Bedrock API key, with the profile ID passed as the model parameter in every case.
Prerequisites include an IAM identity with permission to invoke the profile, an activated source Region (Singapore is enabled by default, while Jakarta, Malaysia, Thailand, and Taipei are opt-in Regions that must be activated first), and a current AWS SDK or CLI. The OpenAI-compatible examples also require the OpenAI SDK and the Amazon Bedrock token generator. For teams migrating an existing OpenAI client, the post advises setting the model parameter on every request, using the supported tool types (function, mcp, custom, namespace, and tool_search), and configuring the inference profile ID directly rather than querying the models endpoint.
The ChatGPT desktop app and the Codex CLI share a single configuration file, so one change covers both. Of the two built-in Amazon Bedrock providers, only amazon-bedrock-runtime accepts an inference profile, and SigV4 signing means there is no API key to rotate. Codex reports a 272,000-token context window for these profiles, which AWS says is Codex’s own bundled figure rather than a service limit; it can be raised with the model_context_window setting. The configuration requires Codex CLI 0.148 or later.
Permissions, Quotas, and Monitoring
Invoking any of the three models through a global profile requires four IAM policy statements: access to the inference profile and the project/default resource in the source Region; access to the foundation model in the source Region scoped to the profile; the Region-agnostic foundation-model ARN with aws:RequestedRegion set to unspecified to permit cross-Region routing; and bedrock:CallWithBearerToken for API-key authentication. Streaming calls add bedrock:InvokeModelWithResponseStream to the first three statements.
A global cross-Region request carries no Region name, so Amazon Bedrock sets aws:RequestedRegion to unspecified and a service control policy with a Region allowlist denies the call, with an error that names only the policy type. The post identifies that as a common reason a first call fails in an enterprise account. Rather than widening the allowlist, AWS recommends scoping the exception with the bedrock:InferenceProfileArn condition key, which Amazon Bedrock populates during authorization; adding unspecified to the allowlist works but opens the same exception to every other service.
For automated offline abuse detection, classifier-flagged traffic is retained by AWS for up to 30 days and processed programmatically. Under cross-Region inference, that retention happens in the destination Region, which for a global profile is any supported commercial Region where the model is deployed. Eligible customers may request full zero data retention through their AWS account team, and the post advises organizations with data residency requirements to involve risk and compliance teams early.
GPT-6 quotas on bedrock-runtime are measured in tokens per minute against the source Region. Input and cache-write tokens count one-to-one, while each output token consumes 10 tokens of quota; the post’s example depletes 12,000 tokens for a request with 2,000 input tokens and 1,000 output tokens. Amazon Bedrock reserves the full max_tokens value up front, and one quota covers InvokeModel, Converse, Responses, and Chat Completions. AWS recommends requesting increases before production and alarming on EstimatedTPMQuotaUsage in the AWS/Bedrock namespace at 70–80 percent of the allocation.
CloudWatch metrics, CloudTrail events, model invocation logs, quota accounting, and billing all stay in the source Region, and customers pay the source Region’s rates rather than the processing Region’s. AWS documents a global profile’s destinations as its commercial Regions rather than enumerating them, and the GetInferenceProfile API returns the Region-agnostic and source-Region model ARNs instead of a destination list. The per-request processing record is additionalEventData.inferenceRegion in CloudTrail in the source Region.
Inference on this endpoint is a management event, logged by default at no extra cost, and CloudTrail Event History retains 90 days, so longer horizons require a trail or model invocation logging. Where developers reach GPT-6 through a coding agent, CloudWatch Coding Agent Insights adds ready-made dashboards under GenAI Observability that break token use down by user, team, and cost center, with the agents emitting OpenTelemetry metrics natively.
