Claude and GDPR: How to Keep Anthropic's Models in the EU

Comparing Claude with other providers? Our comparison of GDPR-compliant LLM APIs sets Anthropic against OpenAI, Azure, Bedrock, Google Cloud, Mistral and four EU clouds. If you also use OpenAI, see our OpenAI and GDPR guide.

This guide explains how each way of reaching Claude handles your data, what Anthropic's DPA and retention terms cover, where the CLOUD Act leaves a gap, and what your organisation must do regardless of which route you pick.

Five ways to use Claude — and only two stay in the EU

“Claude” is not one product with one compliance posture. The same models are sold through five routes, and the one that matters most for GDPR is who runs the inference: Anthropic or a cloud provider.

Claude in Microsoft Foundry is a sixth option, but it offers a US Data Zone only — no EU data zone.

Consumer Claude is a different product

Claude Free, Pro and Max are consumer plans governed by Anthropic's consumer terms, not the DPA. Whether those chats are used to improve models depends on each user's own privacy setting. Employees pasting customer, patient or employee data into a personal Claude account is the same problem as consumer ChatGPT: there is no processor contract covering it. Claude Team, Enterprise and the API fall under Anthropic's commercial terms.

The Claude API directly

Data residency: global or US, nothing else

Two settings control where your data goes. The inference_geo parameter sets where a request is processed, and accepts only global (the default, which may run in any geography) and us. Workspace geo sets where data is stored at rest, and us is currently the only value. US-only inference costs 1.1× the standard rate on current models.

response = client.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=1024,
    inference_geo="us",   # or "global" — there is no "eu"
    messages=[{"role": "user", "content": prompt}],
)
print(response.usage.inference_geo)  # where inference actually ran

For an EU company, us is not an improvement on global: both mean personal data leaves the EEA, and the transfer must be covered by the DPA's Standard Contractual Clauses.

The Data Processing Addendum

Anthropic's DPA is incorporated by reference into its Commercial Terms, so it applies when you accept them — there is no separate signature step. It names you as controller and Anthropic as processor, incorporates the Standard Contractual Clauses (Modules Two and Three) under Irish law, and gives general authorisation for the subprocessors on Anthropic's public list. EEA customers contract with Anthropic Ireland, Limited.

Retention and training

• —Inputs and outputs are deleted from Anthropic's backend within 30 days by default.
• —If the automated trust and safety systems flag a request, inputs and outputs can be kept for up to 2 years, and classification scores for up to 7 years.
• —Retained data is never used for model training without your express permission.
• —Zero data retention (ZDR) is arranged through Anthropic's sales team, per organisation. It is not a console setting.
• —Some of the newest models — currently Claude Fable 5 and 5.1 and Claude Mythos 5 and 5.1 — require 30-day retention and are not available under ZDR unless Anthropic expressly authorises it.
• —Stateful features such as the Batch API (29 days), the Files API and code execution store data by design and sit outside ZDR even if your organisation has it.

Claude on Amazon Bedrock

On Bedrock, AWS runs the models in its own deployment accounts. AWS is the data processor, and model providers have no access to those accounts — so Anthropic never sees your prompts or completions. Bedrock does not store inputs or outputs by default. For most EU companies, this is the cleanest way to use Claude.

Use the eu. inference profile

Current Claude models are served through cross-region inference profiles rather than a single region. The eu. profile keeps processing inside EU regions: called from Frankfurt, it can route to Frankfurt, Stockholm, Milan, Spain, Ireland and Paris. For newer models such as Claude Opus 5 there is no single-region option in Europe, so the EU profile is the only EU route.

import boto3

client = boto3.client("bedrock-runtime", region_name="eu-central-1")

response = client.converse(
    modelId="eu.anthropic.claude-opus-5",   # not global.anthropic...
    messages=[{"role": "user", "content": [{"text": prompt}]}],
)

• —Avoid global profiles. A global. profile routes to every commercial AWS region and is cheaper, which is exactly why teams pick it by accident. You can block it organisation-wide with a service control policy.
• —Check the source region. The destinations depend on where you call from. Calling the EU profile from London adds London; calling from Zurich adds Zurich. Both countries have EU adequacy decisions, but neither is in the EU. AWS advises calling GetInferenceProfile from every source region you use rather than trusting the prefix.
• —Know the retention exceptions. For Claude Fable 5 and 5.1, all traffic is retained for up to 30 days for automated abuse detection, stored by AWS in the region that processed the request. Eligible Enterprise Frontier Safeguards customers get ZDR for these models until 31 December 2026.

Claude on Google Cloud (Google managed)

The Claude models in Google Cloud's Model Garden run on Google's Agent Platform — the product formerly called Vertex AI. Google operates the inference and is the data processor.

• —The EU multi-region endpoint (aiplatform.eu.rep.googleapis.com) keeps processing within the EU. The global endpoint gives no data residency guarantee.
• —Regional endpoints such as europe-west1 (Belgium) serve Claude Sonnet 4.6 and earlier only. For newer models, the EU multi-region endpoint is the EU option.
• —Multi-region and regional endpoints cost 10% more than global.
• —Data handling follows Google Cloud's data governance terms. Request-response logging is off unless you turn it on — and if you do, those logs are personal data you now hold.

from anthropic import AnthropicVertex

client = AnthropicVertex(project_id="my-project", region="eu")  # not "global"

response = client.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": prompt}],
)

The CLOUD Act: the gap no region setting closes

Every route above ends with a US company: Anthropic, Amazon or Google. The US CLOUD Act lets US authorities compel US-controlled providers to produce data they hold, regardless of where it is physically stored. An EU endpoint controls where processing happens; it does not change who can be ordered to hand data over.

For most use cases this is a documented residual risk, not a blocker: record it in your transfer impact assessment and DPIA, minimise what you send, and scrub personal data where you can. For workloads where US jurisdiction is unacceptable, the answer is an EU-headquartered provider rather than a better region — see the EU providers in our LLM API comparison.

What you still have to do yourself

Whichever route you choose, these stay with you as the controller:

• —Pick a route with EU inference (Bedrock eu. profiles or the Google Cloud EU endpoint) if your data must stay in the EU, and enforce it — a service control policy on AWS, a fixed region in code on Google Cloud.
• —Review the DPA and the subprocessor list, and subscribe to subprocessor change notices.
• —Establish a lawful basis for each processing purpose that involves Claude.
• —Run a DPIA where the processing is likely to be high-risk — profiling, employment decisions, health data.
• —Keep personal data out of prompts where you can. Scrub PII before it leaves your infrastructure.
• —Add the processing to your records of processing activities and update your privacy notice.
• —Keep your own logs of AI decisions, stored in the EU. For high-risk AI systems this is a legal requirement under Article 12 of the EU AI Act, and Anthropic's 30-day deletion means you cannot rely on the provider to hold them.

Frequently asked questions

Sources

Provider terms change often. Check the current documentation before you sign or configure anything.

• Anthropic — Data residency
• Anthropic — API and data retention
• Anthropic — How long do you store my organization's data?
• Anthropic — Data Processing Addendum
• Anthropic — Claude Platform on AWS
• Anthropic — Claude Models on Google Cloud
• AWS — Amazon Bedrock data protection
• AWS — Amazon Bedrock abuse detection
• AWS — Supported Regions and models for inference profiles
• Google Cloud — Generative AI data residency