AWS turns on cross-Region routing for GPT-5.6 across 25+ Regions
AWS has switched on cross-Region inference for OpenAI’s GPT-5.6 models on Amazon Bedrock, and the family now runs in more than 25 AWS Regions. AWS posted the service announcement on 17 August 2026 and published the technical walkthrough on 20 August. Three general-purpose variants are covered: Sol, Terra and Luna.
The mechanism is an inference profile, a logical identifier you pass instead of a raw model ID. You call the profile from a source Region, and Amazon Bedrock routes the request to a destination Region using compute in that Region. AWS describes cross-Region inference, or CRIS, as “primarily a capacity mechanism”, which means requests draw on a wider pool of compute rather than one Region’s spare headroom.
All three variants take text and image input, return text, and carry a 1 million token context window. They support reasoning mode, server-side tool calling and prompt caching. You can call them through the OpenAI Responses API, the OpenAI Chat Completions API, or the Amazon Bedrock Converse API.
Two profile types split the routing. A geographic profile, prefixed with a geography code such as us.openai.gpt-5.6-terra, keeps processing inside one geography. A global profile, prefixed global., routes to any supported commercial Region based on real-time capacity. AWS’s cross-Region inference documentation puts global routing at roughly 10% savings against standard pricing, and states there’s no additional routing cost for CRIS at all.
| Geography | Source Regions listed | Examples |
|---|---|---|
| United States | 4 | us-east-1, us-west-2 |
| Canada | 1 | ca-central-1 |
| Europe | 8 | eu-west-1, eu-central-2 |
| Asia Pacific | 12 | ap-northeast-1, ap-south-2 |
| Middle East | 2 | me-central-1, il-central-1 |
| South America | 1 | sa-east-1 |
The catch is quota. Output tokens burn quota at 10x on GPT-5.6, so a single output token consumes 10 tokens from your tokens-per-minute allocation. AWS works the example itself: a request with 2,000 input tokens and 1,000 output tokens depletes 12,000 tokens. Cache read tokens sit outside that calculation, and each cache breakpoint needs a stable prefix of at least 1,024 tokens.
That matters because agent workloads are output-heavy. Together AI’s DeepSWE comparison measured GPT-5.6 Sol at 72.7% pass@1, with a median of 59,000 output tokens per rollout at $8.37 per rollout. Applying the 10x burndown, our calculation puts one of those rollouts at roughly 590,000 tokens against a per-minute quota. Anyone sizing an AI feature’s running cost needs both numbers, not just the billed one.
CRIS is primarily a capacity mechanism.
AWS, Introducing cross-Region inference for OpenAI GPT-5.6 models on Amazon Bedrock
Data residency is where the two profiles genuinely diverge. Data processed through global CRIS may cross any Region in that model’s eligible set, so AWS points workloads with residency rules at a geographic profile or a direct in-Region call instead. Dataconomy reported that AWS also added India geographic profiles for Terra and Luna, routing within Mumbai and Hyderabad. The documentation says cross-Region traffic stays on the AWS network and is encrypted between Regions.
Security and logging don’t change shape. IAM policies control which profiles a role can invoke, Bedrock enforces a zero-operator access model at the chip, and content flagged by the automated abuse-detection classifiers is retained for up to 30 days. For the OpenAI-compatible endpoints, AWS points production users at the aws-bedrock-token-generator package, which mints bearer tokens that expire after at most 12 hours. Requests appear in CloudTrail in your source Region, with additionalEventData.inferenceRegion recording which Region processed each one.
The thing worth watching is the quota split. Geographic and global profiles carry separate allocations, so a load test against one tells you nothing about where the other throttles. AWS says inference profiles don’t support Provisioned Throughput either, which leaves the shared capacity pool as the only lever. For teams already comparing GPT-5.6 Sol on price, that pool is now the number to test.
Get the daily rundown
One email each weekday with the AI news that matters, every claim linked to its primary source.
Free, one email each weekday, unsubscribe in one click. We never sell or share your address.
