Google Vertex AI is Google Cloud's machine-learning platform, and it is where Claude lives for GCP-centric enterprises. As with the other third-party platforms, there is no Anthropic account or API key: your Google Cloud project provides identity, permissions, and billing. This walkthrough gets you from a standard GCP project to a printed Claude response.
Prerequisites: a project, an enabled model, and ADC
A Google Cloud project with billing enabled. Claude usage is metered to a project like any other GCP service, so pick (or create) the project your experiment should be charged to.
Claude enabled in Vertex AI. Inside the Vertex AI section of the console, Anthropic's models are enabled per project. In organizations with a central platform team, this step may require an approval — request it early, since it is the only step you cannot do alone.
Application Default Credentials (ADC). ADC is Google Cloud's standard way for code to find credentials without hardcoding keys. On your laptop, running gcloud auth application-default login once sets it up; in production, a service account attached to the workload provides credentials automatically. If other GCP client libraries already work in your environment, ADC is already working, and the Claude SDK will use it too.
Install the SDK and make the call
Vertex support ships as an extra of Anthropic's official Python package:
pip install -U "anthropic[vertex]"
The client takes two arguments: your project ID and a region. Vertex AI offers a global endpoint, which is the simplest starting point — you can revisit region pinning later if data-location requirements demand it. Note the model ID is the bare first-party form; unlike Bedrock, Vertex does not prefix it.
from anthropic import AnthropicVertex
client = AnthropicVertex(project_id="your-project-id", region="global")
message = client.messages.create(
model="claude-sonnet-5",
max_tokens=512,
messages=[
{"role": "user", "content": "In one sentence, what is Vertex AI?"}
],
)
print(message.content[0].text)If ADC is configured and the model is enabled, this prints a one-sentence answer from Claude Sonnet 5, served through your GCP project. Swap in claude-haiku-4-5 for a faster, cheaper tier or claude-opus-4-8 for the most capable one — the code is otherwise identical.
Log the usage numbers from day one
The response object includes token usage: how many input tokens you sent and how many output tokens Claude generated. These are the units you are billed on, so wire them into your logging from the first script. Pair them with max_tokens — the cap on response length — and you have the two levers that control cost per request before any real cost engineering begins.
What Vertex AI gives you, and what it doesn't
All the core capabilities work here: the Messages API, streaming, tool use, vision, extended and adaptive thinking, and prompt caching. The gaps as of July 2026: the Batch API, Files API, code execution tool, and web fetch tool are not available on Vertex AI, and only a basic variant of the web search tool exists. For chat assistants, summarization, and document workflows, none of that blocks you. If overnight batch processing is central to your plan, factor the gap in now.
The upside is everything GCP-native: Claude calls governed by IAM, visible to your existing audit logging, and billed to the projects and budgets your FinOps process already tracks. The deployment largely inherits your Google Cloud compliance posture — confirm the specifics that matter to your industry with Google.
Where to go next
Go deeper on authentication in Application Default Credentials for Claude on Vertex AI, or check the feature gaps before committing a roadmap. The quickstart shows all four platforms side by side.