The Pulse
AWS Routes GPT-6 Access to Five Asia-Pacific Regions
AWS says Amazon Bedrock users in Thailand, Malaysia, Singapore, Indonesia and Taiwan can access OpenAI’s GPT-6 models through global cross-Region inference. GPT-6 Sol and Luna cover all five source Regions, while GPT-6 Astra is available fr
AI.info Team · Written with AI ·
AWS is extending access to OpenAI’s GPT-6 models on Amazon Bedrock across five Asia-Pacific source Regions, using inference requests that can be processed elsewhere in AWS’s commercial cloud. The company described the setup in a technical post published October 12, covering Thailand, Malaysia, Singapore, Indonesia and Taiwan.
The offering includes GPT-6 Sol and GPT-6 Luna from all five Regions. GPT-6 Astra is available through the global inference profile from Singapore only. AWS’s post is a configuration guide, not an announcement that the models run inside each of those five countries.
Three models, five starting points
Customers invoke the models through Amazon Bedrock Runtime using a global cross-Region inference profile in place of a model ID. For Sol and Luna, the supported source Regions are Singapore, Jakarta, Malaysia, Thailand and Taipei. Astra’s profile lists Singapore as its only source Region.
Bedrock chooses a destination Region where the requested model is deployed. AWS says the destination may be any supported commercial AWS Region, rather than a fixed regional cluster. The arrangement gives calls from the five listed Regions access to a wider pool of inference capacity, but it does not promise that processing stays within the caller’s country.
Existing OpenAI clients can use Bedrock’s endpoint
AWS documents three ways to send requests: the Bedrock Converse API, the OpenAI Responses API and the OpenAI Chat Completions API. The OpenAI-compatible APIs use Bedrock Runtime’s /openai/v1 endpoint and accept either AWS Signature Version 4 or a Bedrock API key. Developers pass the profile ID with each request; AWS also provides setup examples for the ChatGPT desktop app and Codex CLI.
The three models accept text and image input, client-side tool use and prompt caching, according to AWS. Sol and Luna allow reasoning-effort settings from none through max; Astra has the same options except none. AWS describes Sol as suited to reasoning and coding, Luna to high-volume repeatable tasks, and Astra to complex reasoning, coding and long documents.
Cross-Region routing carries residency caveats
Cross-Region inference matters for organizations with strict controls over where requests are processed. AWS says traffic flagged by its automated abuse-detection classifiers can be retained for up to 30 days and processed in the destination Region. Eligible customers may ask their AWS account team about full zero-data retention, and AWS advises organizations with residency obligations to review the arrangement with their risk and compliance teams.
“Your operational surface stays in your source Region.”
Traci Lim, Chris Dickens, Karan Jhaveri, Melanie Li, Saurabh Trikande and Zohreh Norouzi, AWS
AWS says billing, quota tracking, CloudTrail events and CloudWatch metrics remain in the source Region, and customers pay that Region’s rates rather than the destination Region’s. CloudTrail records the processing location in the additionalEventData.inferenceRegion field, giving operators a way to inspect where an individual call ran.
Quota rules can surprise production teams
Bedrock tracks these models’ quotas in tokens per minute against the source Region. Input and cache-write tokens count one-for-one, while each output token consumes 10 tokens of quota. AWS’s example—a request with 2,000 input tokens and 1,000 output tokens—therefore uses 12,000 quota tokens.
The service also reserves the full max_tokens value up front, so AWS recommends setting it close to the expected response length and requesting quota increases before production use. Its guidance suggests alerting when estimated quota use reaches 70% to 80% of the allocation. The profiles make GPT-6 callable from more regional Bedrock endpoints; the processing location, data-retention terms and quota arithmetic remain important parts of that access decision.