GPT-6 lands in five APAC regions on Amazon Bedrock, but four need manual region opt-in

#What happened

Amazon Web Services extended global cross-region inference for all three GPT-6 models to five Asia-Pacific source regions on Amazon Bedrock. The rollout was announced October 12 on the AWS Machine Learning Blog, and Tech Times reported it on October 14, 2026.

GPT-6 Sol and GPT-6 Luna are now callable from Singapore, Jakarta, Kuala Lumpur, Bangkok, and Taipei. GPT-6 Astra is limited to Singapore as a source region. The report calls this the most significant geographic expansion of frontier OpenAI model access in APAC to date.

Only Singapore (ap-southeast-1) is active by default in AWS accounts. Jakarta (ap-southeast-3), Malaysia (ap-southeast-5), Thailand (ap-southeast-7), and Taipei (ap-east-2) are opt-in regions that require explicit account-level activation before Bedrock responds. Teams that skip the step and invoke a global inference profile like global.openai.gpt-6-sol from Bangkok or Taipei will get an OptInRequired error, not a capacity or permissions error.

Access comes through three API paths on the Bedrock runtime endpoint: the Bedrock Converse API, and the OpenAI Responses and Chat Completions APIs at /openai/v1 paths. Teams with existing OpenAI SDK integrations can migrate with a base URL and authentication swap. The ChatGPT desktop app and Codex CLI can also route through Bedrock's APAC regions, and Codex CLI needs version 0.148 or later.

Quota works differently than direct OpenAI billing. Bedrock measures TPM on the bedrock-runtime endpoint, and each output token consumes ten units of TPM quota while each input token consumes one. Bedrock also reserves the full max_tokens value at the start of each request, so a generous max_tokens ceiling can exhaust quota long before actual output volume would. Cached prompt prefixes are excluded from TPM accounting entirely.

#Why it matters

For APAC teams this removes a real capacity ceiling. Global cross-region inference profiles route requests to whichever supported region has headroom, which lowers the chance of throttling under bursty demand, and billing, quota, and CloudTrail logging stay anchored to the source region.

The opt-in step is the gotcha that will waste an afternoon. A team can provision IAM roles, request Bedrock model access, and configure inference profile IDs correctly, and still see errors that look like permissions problems, all because the region itself was never activated.

One caveat for regulated industries. Global profiles can route compute to any supported commercial region, so flagged traffic retained for automated abuse detection may temporarily sit outside the source market. AWS recommends geographic inference profiles or single-region inference when data residency must be guaranteed.

Verified October 14, 2026. Source: Tech Times reporting on the October 12 AWS Machine Learning Blog announcement.