Claude on AWS Bedrock is the way to run Claude in EU regions: Anthropic's own API only knows "us" and "global" as processing locations, while Amazon Bedrock offers an EU profile for Sonnet 5 and Opus 5 that, according to the model card, keeps data within EU regions, plus the AWS data processing addendum, no training on your content and no retention by default. Two catches sit in the fine print: "Claude in Frankfurt" means "Claude in one of the eight eu regions" in practice, and the newest model, Fable 5.1, runs in Europe only through the global profile, with 30 days of retention for abuse detection.
I am a developer, not a lawyer, and this is not legal advice, just the operator's perspective. I use Bedrock in production myself in some cases and have set it up for clients; I went through the AWS and Anthropic documentation again on 21 September 2026 because I wanted to document, for my traffic light for client data, what the yellow path "Claude on EU servers" actually guarantees and what it does not. The rest of my Claude setup is collected on the Claude overview page.
Why Bedrock at all: Anthropic itself has no EU servers
Anthropic's data residency documentation is short and unambiguous. The inference_geo parameter of the Claude API accepts two values, "global" and "us", and the only storage location for a workspace is "us". There is no EU option there, neither for processing nor for storage. For the Team and Enterprise plans I described the same picture in Claude Enterprise or Team: Cost and GDPR: "US-only inference" is a feature, an EU location is not.
For Europe, Anthropic itself points to the partner platforms. Its regional compliance page names AWS Bedrock, Google Vertex and Microsoft Foundry as the ways "to meet local data residency requirements", and the data residency documentation says that on Bedrock it is not the parameter that decides but the inference profile you choose. If you want Claude to run in the EU, you have to change the provider, not the model. That is the reason this post exists.
What Bedrock is, and who you contract with
Amazon Bedrock is the AWS service through which models from various vendors run via API, Claude included. The models do not sit with Anthropic but in AWS-owned accounts: AWS describes one "Model Deployment Account" per region and vendor, into which the model software is copied and to which the vendor has no access. The consequence is spelled out in the documentation: because the providers have no access to those accounts, they have no access to Bedrock logs, prompts or completions. Anthropic phrases it the same way on its own page: Claude in Bedrock runs "with zero operator access", Anthropic personnel cannot reach the inference infrastructure.
Contractually, you are an AWS customer. Usage runs through your AWS account, charges appear on your AWS bill, in Cost Explorer under the model provider. Data processing is governed by the AWS Service Terms, which include the AWS Data Processing Addendum. On top of that come Anthropic's terms for Bedrock, which you accept when enabling the model through AWS Marketplace; for customers in the EEA, Switzerland and the United Kingdom the contracting party there is Anthropic Ireland Limited, and under those terms Anthropic may not train on your content. How the data is handled technically, though, is governed by Bedrock, not Anthropic: "Data handling for this offering is governed by Amazon Bedrock."
The regions: what "Claude in Frankfurt" really means
Bedrock knows three ways to process a request. In-Region stays in exactly one region. Geo routes within a geography such as EU or US. Global routes worldwide. For the Claude models in Europe it looks like this, as of the model cards on 21 September 2026:
The EU profile is called eu.anthropic.claude-sonnet-5 or eu.anthropic.claude-opus-5, and the model card says of it: "Keeps data within EU regions." The sentence next to it matters more: geo and global profiles can route requests outside the source region and do not provide single-region data residency. If you call from Frankfurt, you get processing in one of the eu regions, not necessarily in Frankfurt. AWS counts London and Zurich among its eu regions, two locations outside the EU. Which destination regions a profile actually has from your source region is on no overview page; it comes from the API call GetInferenceProfile, and AWS states that this list never changes for a geo profile. Which region served an individual request is visible afterwards in the CloudTrail field inferenceRegion.
If you genuinely need a single region, you have two addresses: Stockholm and Ireland, where Sonnet 5 and Opus 5 run through the Mantle endpoint with the bare model ID. Frankfurt is not on that list. And Fable 5.1 drops out of any EU commitment entirely: in all eight eu regions only the global profile is available.
The data path: what AWS guarantees and what is retained anyway
The AWS commitments are documented in the Bedrock FAQ and the abuse detection page. Neither AWS nor the model providers train on inputs or outputs. Your content is not shared with providers. Bedrock works with "zero operator access", no operator of the service can reach model inputs or outputs, and by default with "zero data retention", inputs and outputs are not stored. Whatever is processed and stored sits encrypted in the region where you use Bedrock, TLS 1.2 is required in transit, and your own keys and AWS PrivateLink without a path across the public internet are available.
The exception sits in the same documentation, and it concerns exactly the model everyone wants right now. For Claude Fable 5 and Fable 5.1, all traffic is retained for up to 30 days for automated abuse detection, and requests flagged by a classifier may be reviewed by humans at AWS; the documentation names an exception only for one enterprise programme until the end of 2026. To call the model at all, you have to set the mode aws_review in your account, through the data retention API, with no console UI. The retained data stays with AWS, not with Anthropic, and with cross-region profiles it is stored in the destination region. Nothing changes for the earlier models: Opus 4.8, for example, permits the mode none according to the documentation, in which case nothing is retained regardless of your account setting.
For yellow data that means, in my setup: Sonnet 5 or Opus 5 through the EU profile, not Fable. That is not a statement about model quality, it is a statement about the data path.
DPA, standard contractual clauses and the US parent
Data processing at AWS is not an extra step. The AWS DPA is part of the Service Terms, and the European Commission's standard contractual clauses of June 2021 apply automatically according to AWS. AWS is also certified under the EU-US Data Privacy Framework, as part of the certification of Amazon.com, Inc. You determine the region in which your content is stored. So Bedrock covers the part you get no contract for on a single Pro account, without you signing anything; how that looks on Team is in Claude Team: pricing and when it pays off.
What stays with you is the same as with any tool. You need a legal basis for entering personal data regardless of where it is processed; the German data protection authorities say explicitly in their guidance on AI and data protection, margin number 48, that even entering data into an AI system requires a legal basis. The privacy policy, the record of processing activities and the decision about which data belongs in a prompt at all are not taken off your hands by any provider. How I draw that line is in AI and Privacy: What the AI Gets to See.
One point belongs here honestly: the commercial eu regions are operated by a US corporation. AWS's answer to that is the European Sovereign Cloud, live since 15 January 2026 with a first region in Brandenburg, operated exclusively by EU residents, under its own parent company with three German GmbHs. Except: Claude does not appear there. Bedrock in the Sovereign Cloud was extended on 17 September 2026 with Google's Gemma 4 models, Anthropic is not named in any AWS announcement about the Sovereign Cloud, and the Bedrock region overview does not list the region eusc-de-east-1, as of 21 September 2026. If you want Claude, you take the commercial eu regions today, with DPA, standard contractual clauses and the Data Privacy Framework. If you want the Sovereign Cloud, you will find no Claude there today, going by everything AWS has published.
What Claude on Bedrock costs
Bedrock is usage billing, not a subscription. Anthropic's list prices per million tokens as of 21 September 2026: Sonnet 5 costs 2 dollars input and 10 dollars output, Opus 5 costs 5 and 25 dollars, Haiku 4.5 costs 1 and 5 dollars, Fable 5.1 costs 10 and 50 dollars. According to Anthropic, Bedrock has its own regional pricing and the AWS pricing page is authoritative; regional endpoints there carry a 10 percent premium over the global profile according to Anthropic's documentation, and AWS describes the global profile the other way round as "approximately 10% savings".
As an order of magnitude: one million input tokens and 200,000 output tokens on Sonnet 5 come to 4 dollars at list price, and arithmetically 4.40 dollars with the premium Anthropic names; the actual Bedrock prices per region are only on the AWS pricing page. That is a different kind of cost from the 20 dollars for Pro or the 25 dollars per seat on Team that I compare in Claude Pricing and Plans Explained: it is not predictable, but you only pay for what runs, and the bill comes from AWS, not Anthropic.
Claude Code on Bedrock in the EU
Claude Code can run directly against Bedrock instead of through the subscription. Access is a one-off per AWS account: pick an Anthropic model in the Bedrock model catalog and submit the use case form; according to the documentation access is granted immediately. Fable 5.1, Fable 5, Opus 4.8, Sonnet 5, Opus 4.7 and Haiku 4.5 are open to all Bedrock customers; Opus 5 is not on that list, its access criteria are shown in the Bedrock console. You need AWS credentials, via aws configure, an AWS profile or a Bedrock API key. After that, three environment variables are enough:
export CLAUDE_CODE_USE_BEDROCK=1
export AWS_REGION=eu-central-1
export ANTHROPIC_BEDROCK_REGION_PREFIX=eu
With the prefix eu, Claude Code resolves its default models to the EU profiles, for example eu.anthropic.claude-opus-5; in an eu region that is the default anyway; the variable exists from Claude Code 2.1.224 and, according to the documentation, is a preference, not a guarantee. If you want to pin a specific model, set ANTHROPIC_MODEL to the profile ID as well. The documentation names two limitations: the web search tool is not available on Bedrock, and there is no /logout, because sign-in runs through AWS credentials. Everything else that makes Claude Code what it is, files, folders, skills, stays the same; the basics are in Claude Code: A Guide for Beginners.
What is missing on Bedrock
The new Bedrock integration delivers the Messages API with prompt caching, thinking, tool use and citations. Not supported, according to Anthropic, are among others structured outputs, the Files API and URL sources for images and documents, the server-side tools code execution, web search and web fetch, message batches, server-side skills and the MCP connector, and Claude Managed Agents. If you are building your own application, check this list before the decision, not after.
In practice: yellow via Bedrock, red locally
My traffic light for client data has three levels. Green goes in directly: blog posts, copy, anything without personal data. Yellow is personal data in a normal business context, which needs a DPA, anonymisation, or precisely Claude via Bedrock in the EU. Red is health data, third-party bank data and other special categories, which in my setup mostly run directly through local models, no cloud; what such a setup looks like is in Local AI and Privacy: Cloud or Server?.
I use Bedrock in production myself in some cases and have set it up for clients, wherever processing in the EU is required; my own workforce runs on Claude with a DPA day to day. The price is effort: an AWS account, IAM permissions, the question of single region versus EU profile, and a look at CloudTrail when somebody asks where a request was processed. For green data it is not worth it, for red data it is not enough for me. For the middle it is the option I use myself; Anthropic also names Google Vertex and Microsoft Foundry as EU paths, which I have not looked at.
Frequently asked questions
Is Claude on AWS Bedrock GDPR compliant?
No service is compliant on its own, only a setup is. What is documented is what Bedrock delivers: data processing under the AWS DPA with standard contractual clauses, processing in eu regions through the EU profile, no training, no sharing with Anthropic, no retention by default, encryption in transit and at rest. Legal basis, privacy policy and the selection of data remain with you. For your specific case, that belongs in front of your data protection officer.
Does Claude run in Frankfurt on Bedrock?
From Frankfurt yes, in Frankfurt alone no. Sonnet 5 and Opus 5 run in eu-central-1 only through the EU profile or globally. According to the model card the EU profile keeps data within eu regions, but it can route to other eu regions; the destination list for your source region comes from GetInferenceProfile. A single region for these models exists only in Stockholm and Ireland through the Mantle endpoint.
Does Anthropic see my data when I use Claude through Bedrock?
According to the AWS documentation, no. The models run in AWS accounts to which Anthropic has no access, and even the data retained for Fable 5 and 5.1 stays with AWS according to AWS and is not shared with model providers.
Do I need my own DPA with Anthropic?
For processing on Bedrock, AWS is your processor and the DPA is part of the AWS Service Terms. With Anthropic you accept the terms for Bedrock when enabling the model, in the EEA with Anthropic Ireland Limited. Whether your setup needs anything in addition is a question for your data protection officer.
What does Claude on Bedrock cost compared to the Claude API?
The same token list prices as a starting point, with Bedrock's own regional pricing and, according to Anthropic's documentation, a 10 percent premium for regional endpoints over the global profile. Billing runs through your AWS bill. The AWS pricing page is authoritative, not Anthropic's.
Can I use Claude Code through Bedrock?
Yes, with CLAUDE_CODE_USE_BEDROCK=1, an AWS region and the prefix eu for the EU profiles. The web search tool is missing then, everything else stays.
How to make the call for yourself
Three questions and you know whether Bedrock is your path. Does a client or your data protection officer require processing in the EU? If not, a commercial Claude product with a DPA is enough, see Claude Privacy and GDPR: What Applies. If yes: is "somewhere in the eu regions" enough, or does it have to be a single region? That decides between Frankfurt with the EU profile and Stockholm or Ireland with a single region. And do you need Fable 5.1? Then the EU commitment is not available today, and the choice falls on Sonnet 5 or Opus 5.
What no cloud path offers, special categories without any third-party server, remains the territory of local models. How my workforce of fourteen AI employees works within these limits and which rules sit in their personnel files, I show in my Community.
