Vertex AI explained: what runs in the EU in 2026

TL;DR

Vertex AI is Google Cloud’s managed platform for foundation models and agents. Google now calls it Agent Platform. It serves over 200 models, including Claude, Gemini, Llama and Mistral. Eleven European regions exist, and the eu multi-region endpoint keeps machine learning processing inside EU member states for ten percent more than the global rate.

What is Vertex AI, and what is it called now?

Vertex AI is the Google Cloud service that hosts foundation models and the infrastructure around them: a model catalogue, inference endpoints, training, evaluation and an agent runtime. Google’s own documentation now calls it Agent Platform and describes the product as “an evolution of Vertex AI” that “supports the complete AI lifecycle”. The search term and the product name have come apart.

The name changes page maps every component. The rename is not cosmetic for anyone who writes procurement documents or a record of processing activities, because the service you contracted is now listed under a different label.

Previous name	New name
Vertex AI	Agent Platform
Vertex AI Platform	Gemini Enterprise Agent Platform
Vertex AI API	Gemini Enterprise Agent Platform API
Vertex AI Model Garden	Gemini Enterprise Agent Platform Model Garden
Vertex AI Studio	Agent Studio
Vertex AI Agent Engine	Agent Runtime
Vertex AI Search	Agent Search

What does not change: the API host is still aiplatform.googleapis.com, the SDK class is still AnthropicVertex, and model IDs such as claude-sonnet-5 are untouched. Existing code keeps working. We use the name Vertex AI in this guide because that is what people search for, and we treat both names as the same service.

Which models run on the platform?

Google states that the platform offers “over 200 foundation models” across its own families and a curated partner catalogue. The partner models are sold as managed APIs and billed by Google Cloud, so there is no separate vendor contract and no separate invoice.

Provider	Models listed on the platform
Google	Gemini 4 Argon, Gemini 3.1 Pro, Gemini 3.8 and 3.7 Flash, Gemini 2.5 family, Gemma
Anthropic	Claude Fable 5.1, Opus 5.5 and 5, Sonnet 5.5 and 5, Haiku 5.5 and 4.5
Meta	Llama 4 Maverick, Llama 4 Scout, Llama 3.3
Mistral AI	Mistral Medium 3, Mistral Small 3.1, Codestral 2, Mistral OCR
DeepSeek	V4.1-Flash, V3.2, V3.1, R1-0528, OCR
Qwen, GLM, Moonshot, MiniMax, OpenAI	Qwen 3 family, GLM 5.2, Kimi K2 Thinking, MiniMax M2, gpt-oss-120b and gpt-oss-20b
xAI	Grok 4.7, 4.6, 4.3, 4.20

Two caveats matter more than the length of that list. First, availability is per model and per region, so a model in the catalogue is not automatically a model you can call from Europe. Second, Google sets its own lifecycle dates for partner models, and they can be shorter than the vendor’s own schedule. Claude Sonnet 5 on Agent Platform carries a retirement date of “not sooner than December 24, 2026”, which is earlier than Anthropic’s own commitment on the Claude API. Plan a model migration if you need a product to run unchanged for a year.

Which European regions exist?

Google lists eleven European regions for Agent Platform, but only three of them are relevant to a residency decision, because generative AI model availability is narrower than region availability.

Region ID	Location	In the EU
europe-west1	Belgium (St. Ghislain)	yes
europe-west3	Frankfurt, Germany	yes
europe-west4	Netherlands (Eemshaven)	yes
europe-west9	Paris, France	yes
europe-west8	Milan, Italy	yes
europe-west12	Turin, Italy	yes
europe-southwest1	Madrid, Spain	yes
europe-north1	Finland (Hamina)	yes
europe-central2	Warsaw, Poland	yes
europe-west2	London, United Kingdom	no
europe-west6	Zürich, Switzerland	no

The two outside the EU are the ones to watch. London and Zürich sit in Google’s Europe list but outside the EU political boundary, and a resource locations policy written as “europe” will happily include them.

What does the eu endpoint guarantee, and what does it not?

Agent Platform offers three endpoint types, and the one you pick decides where inference happens. The region value you pass when creating the client is the whole mechanism.

Endpoint type	Region value	Host	Where requests are processed
Global	global	aiplatform.googleapis.com	any region with capacity, worldwide
Multi-region	eu or us	aiplatform.eu.rep.googleapis.com	any region in that geography
Regional	e.g. europe-west1	europe-west1-aiplatform.googleapis.com	only the named region

Google’s data residency page is unusually direct about the EU option: “The European Union multi-region (eu) endpoint strictly covers data residency within EU member states”, and “geographies outside the European Union political boundary, including the United Kingdom and Switzerland, are excluded from this endpoint”. That is a stronger statement than most cloud residency wording, and it is the sentence to quote in a data protection impact assessment.

Three limits belong next to it. Data at rest and machine learning processing are governed separately: data stored at rest in your chosen location stays there regardless of which endpoint you call. The global endpoint provides no residency guarantee at all, which Google states as plainly as the EU promise. And regional endpoints serve Claude Sonnet 4.6 and earlier only, so for current Claude models the EU multi-region is the only residency-preserving option, as our guide to Claude on Vertex AI in Europe sets out in detail.

Switching to the EU endpoint is a one-line change:

from anthropic import AnthropicVertex

client = AnthropicVertex(project_id="MY_PROJECT_ID", region="eu")

message = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=512,
    messages=[{"role": "user", "content": "Summarise this clause."}],
)
print(message.content[0].text)

To get there from an empty project:

1. Enable the Agent Platform API and billing on the project.
2. Grant the calling identity a role that includes aiplatform.endpoints.predict, such as roles/aiplatform.user.
3. Accept the partner model terms in Model Garden once, which routes through Google Cloud Marketplace.
4. Set region="eu" and run a test request against claude-sonnet-5.

What does Vertex AI cost in Europe?

Prices are per million tokens in US dollars, and the EU endpoint costs exactly ten percent more than the global endpoint for every model we checked. Anthropic’s documentation for Claude on Google Cloud states the rule plainly: multi-region and regional endpoints cost 10 percent more than global endpoints.

Model	Global input / output	EU multi-region input / output
Claude Sonnet 5	2.00 / 10.00 USD	2.20 / 11.00 USD
Claude Opus 5	5.00 / 25.00 USD	5.50 / 27.50 USD
Claude Opus 5.5	4.00 / 20.00 USD	4.40 / 22.00 USD
Claude Haiku 5.5	0.10 / 0.50 USD	0.11 / 0.55 USD
Claude Fable 5.1	10.00 / 50.00 USD	11.00 / 55.00 USD
Gemini 4 Argon	2.00 / 10.00 USD	2.20 / 11.00 USD

The same split applies to Google’s own models, where the pricing page labels the rows “global” and “non-global”. Gemini 3.8 Flash runs at 0.75 and 3.75 USD globally against 0.825 and 4.125 USD non-global, on a promotional rate that ends on 31 December 2026 and rises to 1.50 and 7.50 USD standard from 1 January 2027. Budget for that step now rather than in January.

Batch inference halves the input and output rate, and cache hits are an order of magnitude cheaper than fresh input. On Claude Sonnet 5 in the EU, a cache hit costs 0.22 USD per million tokens against 2.20 USD uncached. For a workload with a large stable system prompt, caching moves the bill more than the choice of endpoint does. The quota to plan against is 1,250 queries per minute and 12.5 million input tokens per minute on the multi-region endpoint, against 2,500 queries per minute on the global one.

Vertex AI or Bedrock: what EU residency costs on each

Both clouds charge roughly ten percent for keeping inference inside the EU, and both frame it the opposite way round. Google prices the EU multi-region as a surcharge on the global rate. AWS prices its geographic profiles at standard rates and sells the worldwide-routing global profile as “approximately 10% savings”.

	Vertex AI	Amazon Bedrock
Worldwide routing	global endpoint, base price	global cross-Region profile, roughly 10 percent cheaper
EU-bound processing	eu multi-region, base plus 10 percent	geographic EU profile, standard pricing
Single region for current models	not available	available in Frankfurt and Stockholm
Scope of the EU promise	EU member states only, UK and Switzerland excluded	within the geographic boundary of the profile

Claude Sonnet 5 on the Bedrock global cross-Region profile called from Frankfurt is 2.00 and 10.00 USD, identical to the Vertex global rate. AWS also documents that there is no extra routing charge and that the price follows the region you call from. We did not read the geographic EU profile figure off the pricing page in this run, so take the AWS side of that table as the documented rule rather than a quoted number.

The practical difference is not price, it is granularity. Bedrock still lets you pin current Claude models to a single EU region, which our Amazon Bedrock guide covers, while Vertex AI forces current models onto the multi-region endpoint. If your compliance position requires naming one data centre, Bedrock is the easier story. If it requires naming a jurisdiction, Google’s wording about EU member states is the cleaner one, and the exclusion of the UK and Switzerland is explicit rather than inferred. Teams weighing both against a self-hosted option should read our comparison of sovereign AI in Europe first.

One operational limit applies on either platform: Agent Platform caps request payloads at 30 MB, which you can hit with large documents long before you hit the one million token context window.

FAQ

Is Vertex AI the same thing as Agent Platform?

Yes. Google renamed Vertex AI to Agent Platform, and the umbrella product to Gemini Enterprise Agent Platform. The API host, SDK classes and model IDs did not change. Treat documents that say Vertex AI and documents that say Agent Platform as referring to one service.

Does the eu endpoint keep my data in the European Union?

For machine learning processing, yes. Google writes that the EU multi-region endpoint strictly covers data residency within EU member states and that the United Kingdom and Switzerland are excluded. Data at rest is governed separately and stays in the location you selected for storage.

What does Vertex AI cost compared with the global endpoint?

Ten percent more. Claude Sonnet 5 is 2.00 and 10.00 USD per million input and output tokens on the global endpoint, and 2.20 and 11.00 USD on the EU multi-region. The same ten percent applies to Google’s own models and to the other partner models we checked.

Which Claude models can I run in a single European region?

None of the current ones. Regional endpoints serve Claude Sonnet 4.6 and earlier only. For Sonnet 5, Opus 5, Opus 5.5, Haiku 5.5 and Fable 5.1 you use the eu multi-region or the global endpoint.

Vertex AI or Bedrock for a German company?

Bedrock if your requirement names a single data centre, because it still offers in-region inference for current Claude models in Frankfurt. Vertex AI if your requirement names the European Union as a jurisdiction, because Google’s exclusion of the UK and Switzerland is written into the residency documentation. Prices are within ten percent of each other either way.

Do Google’s retirement dates match Anthropic’s?

No, and the gap runs against you. Google sets its own lifecycle dates for partner models. Claude Sonnet 5 on Agent Platform shows a retirement date of not sooner than 24 December 2026. Check the Google model page rather than the vendor schedule before you commit to a model for a year.

Sources

1. Google Cloud: Agent Platform overview (1 November 2026)
2. Google Cloud: Vertex AI name changes (1 November 2026)
3. Google Cloud: Generative AI locations (1 November 2026)
4. Google Cloud: Agent Platform locations (1 November 2026)
5. Google Cloud: Data residency (1 November 2026)
6. Google Cloud: Models on Agent Platform (1 November 2026)
7. Google Cloud: Partner models pricing (1 November 2026)
8. Google Cloud: Agent Platform pricing (1 November 2026)
9. Google Cloud: Claude Sonnet 5 on Agent Platform (1 November 2026)
10. Anthropic: Claude models on Google Cloud (Google managed) (1 November 2026)
11. AWS: Cross-Region inference (1 November 2026)
12. Amazon Bedrock pricing (1 November 2026)