Vertex AI is Google Cloud's machine-learning platform, and Claude is available on it as a third-party model. For an enterprise, the pitch is structural rather than technical: Claude usage lives inside a GCP project, which means it inherits the boundaries, billing, identity, and audit surfaces your organization already governs. There is no new vendor to onboard, no new credential system to secure, and no new invoice to route.
Everything hangs off the project
In GCP, the project is the unit of ownership: it holds the billing account linkage, the IAM policies, the quotas, and the audit logs. Claude on Vertex follows the same rule. Before first use, the Claude models are enabled for the project, and from then on every request is attributed to that project. This makes cost attribution unusually clean — if you give each product team its own project, the bill separates itself. It also makes environment separation natural: dev, staging, and production projects with different access policies and different quota headroom.
Authentication: ADC, not API keys
Vertex does not issue Claude-specific API keys. Requests authenticate with Google Cloud's standard mechanism, Application Default Credentials (ADC) — a discovery convention where the client library automatically finds the right credential for its environment: a developer's logged-in identity on a laptop, or the attached service account on a production workload. The practical consequence is that well-configured code contains no secrets at all.
This is the single biggest "shape difference" for teams arriving from API-key providers, and it is a security upgrade once understood. The mechanics for both developers and services are covered in Application Default Credentials for Claude on Vertex AI.
The SDK, model IDs, and the global endpoint
The official Python SDK ships a Vertex-specific client; install it with pip install -U "anthropic[vertex]". Model IDs are the bare first-party form — claude-sonnet-5, claude-opus-4-8, claude-haiku-4-5 — with no prefix.
from anthropic import AnthropicVertex
client = AnthropicVertex(project_id="my-gcp-project", region="global")
message = client.messages.create(
model="claude-sonnet-5",
max_tokens=1024,
messages=[{"role": "user",
"content": "Draft a polite reply declining this vendor pitch."}],
)
print(message.content)Note the region="global" parameter. Vertex offers a global endpoint that routes requests to available capacity rather than pinning them to a single region — a sensible default for availability. If your data-residency requirements demand a specific processing location, choose a specific region instead and confirm the residency details with Google; Regions, the Global Endpoint, and Data Location Basics explains the trade-off.
Governance the GCP-native way
Because Claude on Vertex is a GCP service, your existing governance stack applies: IAM controls who can invoke models, organization-level policies constrain what projects can do, audit logging records usage, and budgets and alerts watch the spend. For security teams this is the decisive argument — the review reuses conclusions they have already reached about GCP rather than starting from zero. As always with compliance, the platform inherits your cloud provider's compliance posture; confirm the specific certifications and scope with Google rather than assuming.
Know the gaps before you commit
The core feature set — Messages API, streaming, tool use, vision, adaptive thinking, prompt caching — is fully available on Vertex. But as of July 2026, the Batch API, Files API, code execution tool, and web fetch tool are not available, and the web search tool exists only in a basic variant. Managed Agents are not offered (they are exclusive to the first-party API and Claude Platform on AWS). Most document, support, and content workloads will not care; teams with batch-heavy or agentic ambitions should read The Feature Gaps first and weigh whether a second surface belongs in the plan.
Where to go next
Get hands-on with Your First Claude Call on Google Vertex AI, then make authentication production-ready with the ADC guide. For the four-platform comparison, see the platforms section of the main guide.