AWS lets Malaysia-based teams call OpenAI GPT-6 models, with processing outside Malaysia

Amazon Web Services published a technical guide on 12 October 2026 for using OpenAI models through global cross-Region inference. It covers customers calling Amazon Bedrock from Singapore, Jakarta, Malaysia, Thailand and Taipei. AWS says one Region can throttle when demand exceeds its local capacity.[1]

GPT-6 Sol and GPT-6 Luna are available from all five source Regions. GPT-6 Astra is available only from Asia Pacific (Singapore). The models were already on Bedrock; the new post explains the wider routing option and the permissions needed to use it.[1][2][3]

The guide describes three OpenAI models. Sol is aimed at complex work done through the week, Luna at focused, repeatable tasks at high volume, and Astra at complex reasoning, coding and long documents. AWS says all three accept text and image input, client-side tool use and prompt caching.[1]

A developer passes a global inference profile ID in place of a model ID, for example global.openai.gpt-6-sol. The source Region is where the call is made. Bedrock then selects the destination Region that processes it. AWS lists the profiles for Sol and Luna under all five source Regions and the Astra profile under Singapore only.[1]

The earlier AWS posts described general availability. The Sol and Luna post of 22 September said both models cost significantly less than their GPT-5.6 predecessors. The Astra post of 8 September described it as the most capable of the three.[2][3]

AWS says billing, quota tracking, CloudTrail and CloudWatch all remain in the source Region, and customers pay that Region’s rates. Only the source Region needs activating. Singapore is on by default, while Malaysia, Jakarta, Thailand and Taipei are opt-in Regions that an account must enable first.[1]

Processing is different. AWS does not list the destination Regions for a global profile. It describes them as the supported commercial Regions where the model runs. Customers can check where a given request went by reading the inferenceRegion field in CloudTrail, which keeps 90 days of event history by default.[1]

The post’s sample CloudTrail record shows a destination of us-east-1. That is an example, not a statement about where Malaysian traffic will go. It shows that a request made in Singapore can be processed on another continent.[1]

Quotas are counted in tokens per minute against the source Region. Input tokens count one to one, while each output token uses ten units of quota. AWS gives an example of 2,000 input and 1,000 output tokens consuming 12,000 units. Bedrock also reserves the full maximum output size up front, so teams should set it close to what they expect.[1]

For abuse detection, AWS says classifier-flagged traffic is kept for up to 30 days, and under cross-Region inference that retention happens in the destination Region. Eligible customers can ask their AWS account team about full zero data retention. AWS tells readers with data residency requirements to involve risk and compliance early.[1]

Access needs an IAM policy that allows the Region-agnostic model ARN with the requested Region set to unspecified. Organisations that restrict Regions through service control policies may need an exception. The guide supplies an example policy, and AWS recommends readers obtain their own compliance advice.[1]