By Siva Sadhu. Published 28 September 2026. Last checked 29 September 2026.
The short answer
Yes. With a supported Claude model on Amazon Bedrock, the right configuration and calls made from an EU Region, your prompts and responses can be processed entirely within EU Regions. The catch is in the words “supported” and “configuration”. Not every model supports EU routing, and not every “Claude on AWS” path offers it.
Three ways Bedrock routes a request
On Amazon Bedrock, routing is an explicit architecture decision. You make it by choosing what you invoke.
In-Region. Inference is processed within the AWS Region you call, rather than through a cross-Region inference profile. Choose it where your policy, contract or architecture requires single-Region processing. Only some models offer In-Region in EU Regions.
EU Geographic. You invoke an inference profile with the eu.
prefix. Bedrock routes the request among the destination Regions defined
by that profile, and the list depends on the Region you call from.
Called from an EU member-state Region such as Frankfurt, Ireland or
Paris, the profile routes among EU Regions. Called from Zurich, AWS’s
own example shows it can also process requests in Zurich, which is
outside the EU. London is also outside the EU. Choose EU Geographic when
the requirement is “within the EU” rather than “within one Region”, call
it from an EU Region, and confirm the destination list with the
get-inference-profile command before approving it.
Global. A global. profile lets AWS route the request to supported
commercial Regions worldwide. It offers the most capacity and, for some
models, lower prices. Don’t use it where your contract, policy or
approved compliance architecture requires processing within a defined
geography.
Why the model, Region and profile matter
A model can support EU Geographic inference without supporting In-Region inference in every EU Region. Another may only be offered through Global routing from EU Regions. “Available from a Region” doesn’t mean “processed only in that Region”. Check the current Claude EU availability table before approving a model.
For newer models the endpoint matters. Claude Sonnet 5 supports EU Geographic inference on bedrock-runtime and single-Region inference through bedrock-mantle in Stockholm and Ireland. Opus 5.5 also has an EU Geographic profile. Sonnet 5.5 launched on 28 September 2026, but as of this check it does not yet offer EU Geographic or EU In-Region processing on Bedrock. Check the current model card before approving any model.
The path matters too. “Claude on AWS” can mean two different services, and only one of them offers an EU residency control:
| Question | Claude on Amazon Bedrock | Claude Platform on AWS | Anthropic direct |
|---|---|---|---|
| EU inference processing | Yes, with EU Geographic or In-Region on supported models | No. US or Global inference geography only | No EU-only processing control. Global routing may process traffic in Europe, but it cannot be restricted to Europe; stored commercial-product data remains in the US. |
| Who operates inference | AWS | Anthropic | Anthropic and its infrastructure partners |
| Model provider sees prompts | No, according to AWS | Yes, Anthropic is the processor | Yes, Anthropic is the processor |
| Identity and audit | AWS IAM and CloudTrail | AWS IAM; CloudTrail data events if enabled | Anthropic credentials and logs |
| Billing | AWS | AWS | Anthropic |
The one-line test: if you require AWS-operated inference, AWS as the inference data processor, or AWS-enforced regional or geographic processing controls, use Claude on Amazon Bedrock rather than Claude Platform on AWS.
What “processed within the EU” does and doesn’t mean
It does mean that, when you call from an EU Region, inference runs on AWS infrastructure in EU Regions, operated by AWS, and that CloudTrail can record which Region processed each cross-Region request.
It doesn’t mean single-Region processing (unless you use In-Region), that every surrounding service is also in the EU, or that your GDPR analysis is finished. A processing location and a lawful transfer mechanism are different controls. Lawful basis, processor agreements, retention and data classification still matter, and legal interpretation stays with your counsel and DPO.
Configuring it correctly
The important architectural decision isn’t the editor or SDK. It’s which model endpoint and inference profile developers are allowed to invoke. Claude Code can use Bedrock as its backend, but setting only the Region doesn’t pin an EU profile:
{
"env": {
"CLAUDE_CODE_USE_BEDROCK": "1",
"AWS_REGION": "eu-west-1",
"ANTHROPIC_MODEL": "<approved-eu-inference-profile-id>"
}
}
Applications can use Anthropic’s Bedrock client or the AWS SDKs. Newer
integrations can use the AnthropicBedrockMantle client, authorised
through the bedrock-mantle:CreateInference IAM action, so your IAM
policies must cover whichever path you adopt.
- IAM: allowlist approved inference profiles. For cross-Region inference, permissions must cover the profile, the source-Region model and every destination-Region model.
- Organizations: SCPs can add an enforcement layer. Region-deny SCPs must allow the profile’s destination Regions.
- Evidence: CloudTrail can record the invoking principal, the receiving Region, the processing Region and the inference profile. Model invocation logging is separate and off by default.
- Network and safety: add AWS PrivateLink for private connectivity and Guardrails where the model supports them.
What it costs
Claude on Amazon Bedrock is billed per token through your AWS account. Charges come through AWS Marketplace and appear in Cost Explorer under the model provider, not under Amazon Bedrock. For some models, Regional or EU routing costs more than Global routing, so compare prices for your chosen model and routing option before you commit. Set up AWS Budgets and cost alerts before rollout, and tag each workload so costs can be traced back to a team or use case.
The limits
- EU Geographic routing isn’t single-Region, and it stays in the EU only when called from an EU Region. Called from Zurich, it can process requests in Zurich.
- Availability differs by model, endpoint and routing mode, and changes often.
- Global routing isn’t a residency control.
- Certain models, such as Fable 5 and 5.1, require a data retention mode in which AWS retains inputs and outputs for up to 30 days, with possible human review, for safety purposes. Check the model-specific terms before approving a model for sensitive or regulated data.
- Some EU destination Regions (such as Milan and Spain) are opt-in Regions. AWS may store prompts and outputs there for abuse detection, even if you haven’t opted in.
- Default quotas in EU Regions can be low. Request increases early, before a production launch.
- The AWS European Sovereign Cloud is a separate partition. Bedrock is available there with open-weight models such as Gemma 4, but Claude isn’t currently part of that offering.
- Model versions have end-of-life dates. Today’s approved model can become tomorrow’s unsupported dependency.
- Server-side Anthropic features such as web search or code execution may not be available through a given Bedrock integration.
Need an EU residency control you can prove?
We help teams design, build and evidence Claude deployments on Amazon Bedrock. Book a consultation.
Sources
- Amazon Bedrock: Regional availability by model
- Amazon Bedrock: Supported cross-Region inference profiles
- Amazon Bedrock: Geographic cross-Region inference
- Amazon Bedrock: Claude Sonnet 5 model card
- AWS Alps blog: Cross-Region inference for EU data processing
- AWS: Claude Platform on AWS data residency
- Anthropic: Claude in Amazon Bedrock
- Anthropic: Data residency
- AWS Security blog: Run open-weight models on Amazon Bedrock in the AWS European Sovereign Cloud
Where vendor documentation differs from this article, the vendor documentation is authoritative.

