In-depth analysis · About 4 min read
Bedrock’s Canadian access to GPT models still requires an explicit inference geography
An AWS guide distinguishes Canadian API endpoints from US geographic and global routing, including the destination-region abuse-review exception.
- Published here
- Source published
- Source checked
AWS announced on October 20 that customers can access GPT-5.6, GPT-6 and GPT-6.1 variants through Bedrock cross-Region inference from Canada (Central) and Canada West (Calgary). This is a Bedrock access and routing update. Calling a Canadian endpoint does not mean the model necessarily processes the request within Canada, nor does it confer a personal ChatGPT subscription benefit.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- The guide covers GPT-5.6 Sol, Terra and Luna; GPT-6 Astra, Sol and Luna; and GPT-6.1 Sol. US geographic profiles can route within Canada and the US; global profiles can use supported commercial AWS Regions.
- GPT-5.6 global profiles are listed only for ca-central-1; Calgary uses the US geographic profile for those variants. Flagged content may be retained for up to 30 days in the destination Region for abuse review.
The source Region and execution Region are different
An application calls Bedrock Runtime in its source Region and passes an inference-profile identifier instead of a raw model ID. Bedrock routes to a supported destination without application-managed destination selection. The us. profile is a North American boundary covering Canada and the US, not a Canada-only promise. The global. profile draws on supported commercial deployment Regions. The article lists both profile types for GPT-6 and GPT-6.1 from both Canadian sources, but GPT-5.6 global only from Canada Central. Before deployment, inspect active profiles for the actual source Region. Broader capacity is a routing mechanism, not a measured guarantee of throughput or latency for an individual application.
API familiarity does not remove AWS authorization
The guide demonstrates Responses, Chat Completions and Bedrock Converse on the regional Bedrock Runtime endpoint. OpenAI-compatible paths use /openai/v1; requests still rely on AWS SigV4 or a Bedrock API key. Its SDK example derives a short-term token from existing AWS credentials rather than storing a static key. IAM must cover the chosen profile and model resources, with additional bearer-token or Responses project permissions where applicable. Organization service control policies must allow the selected routing path and required destinations. A tool-call output is a structured request for the client application to execute, not proof that a function ran. The model descriptions mention a one-million-token context and tool support; exact modalities and parameters remain model-specific.
Caching and quota are separate accounting questions
AWS says eligible prompt prefixes of at least 1,024 tokens can be cached implicitly. Cached reads receive discounted billing and do not count against the tokens-per-minute quota; the guide points to model-specific rules for write pricing and explicit breakpoints. This does not establish a complete request price. Cross-Region quotas combine input and output, with output weighted by a model-specific burndown rate. US geographic and global profiles have separate allocations, so changing profile changes the quota pool as well as geography. A representative trial should include output length, concurrency, streaming and peak traffic. AISubHub did not run that trial, and no quota increase or caching benefit is assumed available to a particular account.
Audit location is not the same as all data staying local
CloudWatch metrics and invocation logs are recorded in the source Region even when processing occurs elsewhere. CloudTrail’s additionalEventData.inferenceRegion identifies the serving destination. Full request/response invocation logging is off by default and must be explicitly enabled, with its own data-access implications. The guide says ordinary at-rest data stays in the source Region by default and describes encrypted transport over AWS’s backbone. It also identifies an OpenAI-model exception: content flagged by automated abuse classifiers may be retained for up to 30 days in the destination Region for review. Choosing us. constrains processing to Canada and the US, not to Canada alone. This exception must remain visible when evaluating data residency; neither a regional log location nor an API-compatible client removes it.
Why it matters
AISubHub analysis: model selection and processing geography form one deployment decision. Source-region logs and a Canadian client URL cannot substitute for verifying the active profile’s destinations and the organization’s own data-handling requirements.
Limits and availability
We read the AWS guide and did not invoke the models, inspect a customer account, measure latency or verify quotas. Profile membership and availability can change. This is a technical reading, not a conclusion that a particular workload meets legal or regulatory obligations; no price is inferred.
Sources and evidence
Source published: · Source checked:
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.