Claude on Vertex AI in Europe: EU Endpoint Guide 2026

TL;DR

To run Claude on Vertex AI in Europe, set the region to eu, which calls the EU multi-region endpoint aiplatform.eu.rep.googleapis.com. Opus 5, Sonnet 5 and Fable 5.1 are served there, processed only in EU member states, at a 10% premium over the global endpoint. Single-region endpoints such as europe-west1 only cover older models.

Is it still called Vertex AI?

Mostly in people’s heads. Google’s documentation now files Claude under “Gemini Enterprise Agent Platform”, and the old Vertex AI pricing URL redirects there. The API host is still aiplatform.googleapis.com, the Anthropic SDK class is still AnthropicVertex, and every code sample in this guide works unchanged. We keep the Vertex name because that is what people search for.

This matters for two practical reasons. Internal tickets and procurement records often reference “Vertex AI”, while current Google documentation and console menus use the new name. And Anthropic’s own Google Cloud guide now says “Agent Platform”. Treat both names as the same service. Nothing about data processing terms changed with the rebrand, as far as the documentation shows.

Which Claude models are available in Europe on Vertex AI?

Claude Opus 5, Sonnet 5 and Fable 5.1 are available in Europe through the eu multi-region endpoint. Haiku 4.5 is available in the Belgian region europe-west1. None of the three current flagship models can be pinned to a single European region such as Frankfurt or the Netherlands, so plan around the multi-region endpoint.

The details come from Google’s model pages for Sonnet 5, Opus 5 and Haiku 4.5.

Model	Model ID	European location	Default quota in Europe	Google retirement (not sooner than)
Claude Opus 5	claude-opus-5	eu multi-region	1,000 QPM, 10M input TPM	24 January 2027
Claude Sonnet 5	claude-sonnet-5	eu multi-region	1,250 QPM, 12.5M input TPM	24 December 2026
Claude Fable 5.1	claude-fable-5-1	eu multi-region	see Google quota page	1 March 2027
Claude Haiku 4.5	claude-haiku-4-5	europe-west1	1,800 QPM, 1.8M input TPM	15 October 2026

Look closely at the retirement column. Google sets its own dates, and they are earlier than Anthropic’s. Anthropic commits to Sonnet 5 until at least 30 June 2027 on its own API, but Google only commits to 24 December 2026. If you build a product that must run unchanged for a year, factor in a model migration on Vertex.

What does the EU multi-region endpoint guarantee?

The eu endpoint keeps ML processing inside EU member states and stores data at rest in the selected multi-region. Google states explicitly that the United Kingdom and Switzerland are excluded. The global endpoint gives no residency guarantee at all, because Google may process requests in any Google Cloud location worldwide.

Google’s data residency page separates two things: where data rests and where ML processing, meaning model inference, happens. For partner models the multi-region endpoint is the only way to control the second for the newest Claude models.

Endpoint	Host	ML processing	Price vs. global	Claude 5 models
Global	aiplatform.googleapis.com	Anywhere in the world	Baseline	Yes
EU multi-region	aiplatform.eu.rep.googleapis.com	EU member states only	+10%	Yes
Regional europe-west1	europe-west1-aiplatform.googleapis.com	Europe (Haiku 4.5: EU multi-region)	+10%	No

This is stricter than Amazon’s EU profile in one respect: on Bedrock, calling the EU profile from London or Zurich adds those regions as destinations. On Google, the eu endpoint never leaves the EU. For a German, French or Dutch company that is the cleaner story for a data protection impact assessment.

How do you enable Claude on Vertex AI?

You need a project with billing, the Agent Platform API, two IAM roles and a one-time model enablement in Model Garden. An administrator must also allow the Cloud Commerce procurement API in your organisation policy, because Claude is sold through Google Cloud Marketplace. Plan 30 to 60 minutes, most of it waiting for approvals.

The steps follow Google’s partner model guide:

1. Enable the API. Run gcloud services enable aiplatform.googleapis.com --project=PROJECT_ID.
2. Grant the enablement role to whoever clicks “Enable” in Model Garden: roles/consumerprocurement.entitlementManager.
3. Grant the inference role to your application’s service account: roles/aiplatform.user, which contains aiplatform.endpoints.predict.
4. Check organisation policies. Allow cloudcommerceconsumerprocurement.googleapis.com. If you use Assured Workloads, Google notes that you may need an exception for this API and for commerceagreement.googleapis.com.
5. Enable the model on its Model Garden card and accept the Anthropic terms.
6. Block the global endpoint with the organisation policy constraint constraints/gcp.restrictEndpointUsage, so no developer can bypass the EU endpoint by accident.

The role bindings, as single-line commands:

gcloud projects add-iam-policy-binding PROJECT_ID --member="user:admin@example.eu" --role="roles/consumerprocurement.entitlementManager"

gcloud projects add-iam-policy-binding PROJECT_ID --member="serviceAccount:claude-app@PROJECT_ID.iam.gserviceaccount.com" --role="roles/aiplatform.user"

Anthropic resellers are a known blocker. Google warns that if your billing account is managed by a reseller that Anthropic prohibits, you cannot accept the terms or enable Claude. Talk to your reseller before the project kickoff, not after.

What does the Python code look like with anthropic[vertex]?

Install anthropic[vertex], authenticate with Application Default Credentials and create an AnthropicVertex client with region="eu". The SDK then targets the EU multi-region host automatically. The model name goes into the call as usual, and anthropic_version is handled by the SDK for you. Claude Code uses the same EU endpoint with CLOUD_ML_REGION=eu, as shown in our Claude Code Bedrock and Google Cloud setup guide.

pip install -U "anthropic[vertex]"
gcloud auth application-default login

from anthropic import AnthropicVertex

client = AnthropicVertex(project_id="my-eu-project", region="eu")

message = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Summarise GDPR Article 28 in two sentences."}],
)

print(next(block.text for block in message.content if block.type == "text"))
print(message.usage)

For streaming, swap messages.create for messages.stream and iterate over stream.text_stream. For Haiku 4.5 in Belgium, use region="europe-west1" and model="claude-haiku-4-5@20251001", which is the ID Anthropic documents for Google Cloud.

If you prefer plain HTTP, for example to test a firewall rule, this curl call hits the EU endpoint. Note that the model sits in the URL and anthropic_version sits in the body.

curl -X POST "https://aiplatform.eu.rep.googleapis.com/v1/projects/PROJECT_ID/locations/eu/publishers/anthropic/models/claude-sonnet-5:rawPredict" -H "Authorization: Bearer $(gcloud auth print-access-token)" -H "Content-Type: application/json" -d '{"anthropic_version":"vertex-2023-10-16","max_tokens":256,"messages":[{"role":"user","content":"Hello from the EU endpoint"}]}'

What does Claude on Vertex AI cost in Europe?

The EU multi-region carries the same 10% premium as Bedrock’s EU profile. Sonnet 5 costs 2.20 USD per million input tokens and 11 USD per million output tokens, Opus 5 costs 5.50 and 27.50 USD. On the global endpoint the prices match Anthropic’s list prices exactly. Google bills in USD on your Cloud invoice.

Figures from the Google Cloud generative AI pricing page, tabs “Global” and “EU Multi-Region”, read on 18 September 2026:

Model	Global input / output	EU multi-region input / output	EU batch input / output	Example: 10M in + 2M out per month (EU)
Claude Opus 5	5.00 / 25.00	5.50 / 27.50	2.75 / 13.75	110 USD
Claude Sonnet 5	2.00 / 10.00	2.20 / 11.00	1.10 / 5.50	44 USD
Claude Fable 5.1	10.00 / 50.00	11.00 / 55.00	5.50 / 27.50	220 USD
Claude Haiku 4.5	1.00 / 5.00	1.10 / 5.50 (europe-west1)	0.55 / 2.75	22 USD

All prices in USD per million tokens. The EU cache hit price for Sonnet 5 is 0.22 USD, a tenth of the input price, so caching long system prompts is the biggest lever you control. Also remember that Claude 4.7 and later count roughly 30% more tokens for the same text than older models, according to Anthropic’s pricing notes.

Vertex and Bedrock cost the same for the Claude 5 models in the EU. Choose the cloud on residency scope, features and where your data already lives, not on token price. Our Claude API pricing EU comparison adds the Anthropic API and Foundry to the picture.

How does Google handle your prompts?

Google acts as your processor under the Cloud Data Processing Addendum, does not train on your data without permission and does not share prompts or responses with third parties, Anthropic included. The catch is the Fable models: Google stores their prompts and responses for up to 30 days for abuse monitoring, so zero data retention is not possible there.

Google’s zero data retention page lists the conditions:

• Training restriction. Google will not use your data to train or fine-tune any AI or ML model without your prior permission. This applies to all managed models, including partner models.
• Abuse monitoring. Prompt logging for abuse monitoring is described for Google’s own models, and customers can request an exception.
• Advanced AI models. The Fable 5.1 model page states that all prompts and responses are stored for up to 30 days for abuse monitoring under the Advanced AI Safety Addendum. Anthropic lists Fable 5.1 as a “Covered Model” with 30-day retention on every platform.
• Request logging. Anthropic recommends enabling 30-day request-response logging for misuse investigation. That log sits in your project and is your data, so include it in your records of processing.

For Opus 5 and Sonnet 5 on the eu endpoint, our assessment is that Vertex gives European companies one of the cleanest residency setups currently documented for Claude. The AI Act side, including logging duties, is in our EU AI Act checklist for LLM deployers.

Which limits and pitfalls should you plan for?

The main pitfalls are feature gaps against the Anthropic API, the earlier Google retirement dates, private networking on multi-region endpoints and payload size. None of them block a typical chat or RAG workload, but each can break a migration that assumed feature parity with the Anthropic API.

1. Unsupported features. Anthropic lists the Files API, code execution, web fetch, the MCP connector, Agent Skills, the Anthropic Message Batches endpoint and Managed Agents as unsupported on Google Cloud. Web search and structured outputs are supported.
2. Private connectivity. Private Google Access does not work with multi-region endpoints. Google requires Private Service Connect endpoints for regional Google APIs instead, configured with gcloud.
3. Payload limit. Requests are capped at 30 MB. Large PDFs or many images hit that before the 1M token context does.
4. Quota is per endpoint. The eu multi-region has its own quota, half of the global one for Sonnet 5. Request increases early if you expect production traffic.
5. Retirement dates. Plan a migration window for Sonnet 5 on Vertex before 24 December 2026, or keep an abstraction layer that lets you swap the model ID.

Still comparing clouds? Our Claude on AWS Bedrock in the EU guide covers Amazon’s EU profile and in-region options, and the GDPR comparison of Bedrock, Vertex, Foundry and the Anthropic API gives the full decision matrix.

FAQ

Each answer below stands on its own, so you can link to any single question from internal documentation. Everything reflects Google Cloud and Anthropic documentation as read on 18 September 2026. Google updates model availability and quotas frequently, so recheck the model pages before you commit.

Can I run Claude Sonnet 5 in Frankfurt on Vertex AI?

No. Sonnet 5 and Opus 5 are available on Google Cloud only through the global endpoint and the US and EU multi-region endpoints. The EU multi-region keeps processing within EU member states, but you cannot choose Frankfurt (europe-west3) or any other single European region for these models.

What does the EU endpoint cost compared with the global endpoint?

It costs 10% more on every token type. Sonnet 5 is 2.20 USD instead of 2.00 USD per million input tokens and 11 USD instead of 10 USD per million output tokens. A workload of 10 million input and 2 million output tokens a month costs 44 USD on the EU endpoint instead of 40 USD.

Vertex AI vs. AWS Bedrock for Claude in the EU: which is better?

For pure EU residency, Vertex has the stricter scope because its eu endpoint excludes the UK and Switzerland by definition. Bedrock is stronger if you need a single named region, since it offers in-region Opus 5 and Sonnet 5 in Ireland and Stockholm. Token prices for the Claude 5 models are identical.

Is my data used to train Claude or Gemini on Vertex AI?

No. Google’s training restriction says it will not use customer data to train or fine-tune any AI or ML model without prior permission, and this covers partner models such as Claude. Google also states that prompts and responses are not shared with third parties, which includes Anthropic.

Which SDK region value do I use for the EU endpoint?

Use region="eu" in AnthropicVertex. The SDK then calls https://aiplatform.eu.rep.googleapis.com. Use region="global" only if you accept worldwide processing, and use a named region such as europe-west1 only for models that support it, currently Haiku 4.5 and older Claude versions.

Does zero data retention work for Claude Fable 5.1 on Google Cloud?

No. Google’s model page states that all Fable 5.1 prompts and responses are stored for up to 30 days for abuse monitoring, and Anthropic classifies Fable 5.1 as a Covered Model that requires 30-day retention on every platform. For zero retention, use Opus 5 or Sonnet 5 instead.

Sources

1. Anthropic docs: Claude on Google Cloud (18 September 2026)
2. Google Cloud docs: Claude Sonnet 5 (18 September 2026)
3. Google Cloud docs: Claude Opus 5 (18 September 2026)
4. Google Cloud docs: Claude Haiku 4.5 (18 September 2026)
5. Google Cloud docs: Use partner models (18 September 2026)
6. Google Cloud docs: Data residency (18 September 2026)
7. Google Cloud docs: Zero data retention (18 September 2026)
8. Google Cloud generative AI pricing (18 September 2026)