Skip to main content
The gateway serves models from five providers. Every model runs under a signed BAA — AWS or Microsoft. Whatever model you pick, the redaction path is identical: identifiers never reach the model. Where each model processes your data is covered by our PDPA Transfer Impact Assessment, which names every destination. We share it under NDA — ask us, and ask before you build against a model if residency is a requirement for you. The assessment is ours to complete, not yours to negotiate. A model whose destination we have assessed is callable by every customer. One we have not assessed is callable by none, and returns 403 with code: "model_not_allowed" until that changes. The Transfer assessment column below tells you which is which today. If you omit model, the gateway uses apac.amazon.nova-pro-v1:0. Read the model ID as a destination, not just a capability. A prefix names where inference happens: apac. is the ten Asia-Pacific AWS regions, us. is the United States, in. is India, and jp. is Japan. A model ID with no such prefix, or with global., is served from a pool that names no single jurisdiction — which is why those cannot carry a completed assessment, whatever their BAA status. A signed BAA covers the service; a transfer assessment needs the country. Both are required, and they are not the same question. Every model in this catalog supports tool calling. Models without tool support are not listed.

​
Amazon

The default family, and the one with a completed assessment. Pick these unless you have a reason not to.

​
Anthropic

Same AWS account and BAA as the Nova family. The BAA is not the open question — the destination is. A global. profile routes to whichever region has capacity, so there is no single country to assess, and by the rule above that keeps the assessment Open. An APAC-resident Claude exists on a jp. profile and would close it; that route is not served here yet.

​
OpenAI

The us. and in. prefixes are the destination. These are the same GPT-5.6 models, served from a single named jurisdiction — the United States, or India — rather than from a pool. That is what makes their assessment completable, and it is why the prefix is part of the model ID rather than a separate parameter: naming the model names the destination. Pick in. for the shortest chain — one country, one assessment. Pick us. for the widest capacity. The global. rows are the same three models served from a worldwide pool: same BAA, same processor, but no single country to assess — so they stay Open, and clinical traffic belongs on a prefixed id.
The GPT-5.6 family retains flagged traffic for up to 30 days. AWS applies zero data retention to Bedrock by default, but the GPT-5.6 family is an exception: traffic flagged by its abuse classifier is held up to 30 days for automated review, in the region that processed the request. It is not shared with OpenAI. If that does not suit your use case, pick a Nova or gpt-oss model — neither carries the exception.
Dialect note: the AWS-served GPT-5.6 route speaks Bedrock Converse. The gateway routes by model ID, so your request format does not change.

​
DeepSeek

​
MiniMax

Note: this model emits reasoning tokens billed as output, so its cost per task runs above its per-token rate card.

​
Choosing a model

The catalog is the same for every customer. A model is callable once its destination is covered by our transfer assessment, and refused until then — the same set for everyone, not a per-account menu. Models not yet covered return 403 with code: "model_not_allowed".There is no region or routing parameter on the request. The model you name determines how it is served, so the model id is the whole choice — not a per-call flag we would have to trust the caller to set.

​
Errors you may see

See Errors & limits for the full error contract.