Skip to content
NewsProduct launch

AWS routes OpenAI's GPT-6 models on Bedrock across regions for five Southeast Asian markets

· by Pondero Newsdesk

The short version

AWS published a how-to on October 12 letting developers in Thailand, Malaysia, Singapore, Indonesia, and Taiwan call OpenAI's GPT-6 models on Bedrock through cross-Region inference, trading local capacity limits for a quota formula that charges 10x for output tokens.

AWS routes OpenAI's GPT-6 models on Bedrock across regions for five Southeast Asian markets

AWS published a how-to on October 12 that lets developers in Thailand, Malaysia, Singapore, Indonesia, and Taiwan stop waiting on their own region's Bedrock capacity for OpenAI's GPT-6 models, and instead draw on whatever AWS region has room, while every quota check, bill, and audit log still lands at home, per AWS's engineering blog.

What AWS changed

Amazon Bedrock now offers OpenAI's GPT-6 Sol and GPT-6 Luna through global cross-Region inference profiles callable from five source regions: Singapore, Jakarta, Malaysia (ap-southeast-5), Thailand (ap-southeast-7), and Taipei, per AWS. GPT-6 Astra, the most capable of the three, is callable only from Singapore. Developers invoke the models by passing a profile ID such as global.openai.gpt-6-sol in place of a model ID on the Converse API, the OpenAI Responses API, or the OpenAI Chat Completions API; Bedrock then picks whichever supported commercial region actually processes the request. Four of the five source regions (Jakarta, Malaysia, Thailand, Taipei) are opt-in regions a customer has to activate first; Singapore is on by default. Authentication works with either AWS Signature Version 4 or a Bedrock API key.

Billing stays home, but the quota math gets stricter

The post's most consequential detail is not the new regions but how AWS counts against them. Token quotas are tracked per source region in tokens per minute, and input and cache-write tokens count 1:1 against that quota, but every output token consumes 10 units, so a call with 2,000 input tokens and 1,000 output tokens burns 12,000 tokens of quota, per AWS's worked example. Bedrock also reserves a request's full max_tokens setting up front rather than metering actual output after the fact, so a generous cap eats quota whether or not the model uses it. Teams that were already running close to their regional ceiling and assumed cross-Region inference would simply unlock more headroom need to re-run their quota math first, tighten max_tokens, and set a CloudWatch alarm on EstimatedTPMQuotaUsage before they lean on the new routing in production. The payoff is real: billing, CloudTrail events, and invocation logs all stay in the calling region, so a team does not have to re-architect its compliance posture to get access to capacity elsewhere.

What to watch next

Astra's single-region limitation to Singapore is the gap most likely to move next; AWS's post frames capacity, not geography, as the constraint, so an expansion to more Astra source regions would follow actual deployment, not a policy change. Developers should also watch AWS's token-counting documentation for updated quota figures, since the post points to Bedrock's quotas page for current values rather than quoting a fixed ceiling.

Sources