AWS Expands OpenAI GPT-6 Global Inference Across Five Asia Pacific Regions

AWS opened five Asia Pacific source Regions on October 12

Amazon Web Services announced on October 12, 2026 that customers can invoke OpenAI GPT-6 Sol and GPT-6 Luna through Amazon Bedrock global cross-Region inference from Asia Pacific (Singapore), Jakarta, Malaysia, Thailand and Taipei. GPT-6 Astra is available through the same global route from Singapore. The exact profile IDs are global.openai.gpt-6-sol, global.openai.gpt-6-luna and global.openai.gpt-6-astra.

This is an access and capacity expansion, not a new OpenAI model release. Singapore uses ap-southeast-1; Jakarta ap-southeast-3; Malaysia ap-southeast-5; Thailand ap-southeast-7; and Taipei ap-east-2. AWS says the latter four are opt-in Regions that an account must activate before use.

Global routing broadens capacity and the data boundary

A request starts at bedrock-runtime in the selected source Region, but Amazon Bedrock chooses a supported commercial AWS Region where the model is deployed to perform inference. Billing, token quota, CloudWatch metrics, invocation logs and CloudTrail records remain in the source Region. The actual processing Region appears per request as additionalEventData.inferenceRegion in CloudTrail.

That split matters: a source endpoint in Thailand, Malaysia, Singapore, Indonesia or Taiwan does not establish that model processing stayed there. AWS does not enumerate one fixed destination list for the global profile. Teams with country, geography or single-Region requirements should use an explicitly compatible local or geographic route, or wait; a global profile is the wrong control boundary.

The three models share a large interface but not a price

The live model cards list a 1,050,000-token context window and 128,000 maximum output tokens for GPT-6 Sol and Luna, and the reviewed Astra card documents the same 272,000-token short-versus-long-context pricing boundary. The models accept text and image inputs and return text, including code. Sol and Luna accept reasoning effort from none through max; Astra accepts low through max but not none.

AWS positions Luna for focused repeatable work, Sol for demanding development and multi-step work, and Astra for complex reasoning, coding and long documents. Those are provider descriptions, not comparative evidence. Buyers should evaluate the exact model, reasoning setting, API and workload rather than infer quality from the tier name.

Short-context Global CRIS pricing spans 100×

For requests with at most 272,000 input tokens, the live Global CRIS Standard rates per million tokens are: Luna $0.10 input, $0.125 cache write, $0.01 cache read and $0.50 output; Sol $2 input, $2.50 cache write, $0.20 cache read and $10 output; Astra $10 input, $12.50 cache write, $1 cache read and $50 output. AWS says Global CRIS carries no regional premium and bills at the source Region’s applicable rate.

Above 272,000 input tokens, higher prices apply to the entire request: Luna $0.20 input and $0.75 output, Sol $4 input and $15 output, and Astra Standard $20 input and $75 output per million tokens, with separate cache rates. Astra Ultrafast is six times its Standard price, reaching $60 input and $300 output for short-context Global CRIS. Prices can change, and quality or latency parity is not established by a compatible endpoint.

OpenAI-compatible APIs reduce—not remove—migration work

The bedrock-runtime endpoint exposes the Bedrock Converse and InvokeModel APIs plus OpenAI-compatible Responses and Chat Completions paths under /openai/v1. A client must pass the global inference-profile ID as model. AWS documents Signature Version 4, Bedrock API keys and short-lived bearer-token generation from AWS credentials, depending on the client path.

Compatibility is not equivalence. AWS tells migrating clients to set the model on every request, configure the profile directly rather than querying the models endpoint, and use supported tool types. Teams still need contract tests for request fields, tool calls, structured output, streaming, reasoning blocks, errors, retries, cache accounting and response persistence before moving production traffic.

Retention can follow the destination Region

AWS describes Bedrock as zero operator access and zero data retention by default, but its abuse-detection page names GPT-6 Astra, Sol and Luna as exceptions: classifier-flagged traffic may be retained for up to 30 days for automated offline abuse detection. AWS says retained content is not shared with the model provider.

For cross-Region inference, retained inputs and outputs are stored in the destination Region where processing occurred. Eligible customers may request full zero-data-retention access through their AWS account team, but eligibility is not automatic. Treat that approval and the effective account configuration as required evidence—not an assumption based on the general Bedrock description.

Permissions and quota behavior can surprise an existing account

AWS documents permissions for the source-Region inference profile, the source-Region foundation model and a Region-agnostic foundation-model ARN. Global calls use aws:RequestedRegion set to unspecified, so an organization SCP that only allows named Regions can deny the request. AWS recommends a profile-scoped exception instead of adding unspecified broadly to a Region allowlist.

The runtime quota is measured in tokens per minute in the source Region. Input and cache-write tokens count one-for-one, while each output token consumes ten quota tokens; Bedrock also reserves max_tokens up front. A 2,000-input, 1,000-output request therefore depletes 12,000 quota tokens. Production sizing must use this quota arithmetic rather than the billed-token total.

What builders and buyers should do next

Trial Global CRIS when broader commercial-Region routing is acceptable and capacity matters. Start with non-sensitive traffic, the least expensive model that clears a frozen workload evaluation, and an explicit source Region. Capture account-visible access, current prices, effective retention mode, quota, profile ARN, API behavior and CloudTrail inferenceRegion records.

Constrain rollout with profile-scoped IAM and SCP rules, maximum output limits, spend and quota alarms, destination monitoring, a tested fallback and a prohibition on silent route expansion. Wait when ZDR approval, destination scope, model access or a price quote is unresolved. Reject this route when processing or retained data must remain in the source country, geography or Region.

Disclosures

AccessAllGPT received no AWS or OpenAI account, credentials, credits, access, pricing, traffic data, briefing, review, payment or compensation for this article. AWS and OpenAI did not sponsor, review or endorse it. AccessAllGPT did not test, score or rank the models or service. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with AWS, OpenAI or organizations cited. Publication-wide relationships are listed on the disclosures page.