Enabling Claude in Vertex AI Model Garden

On Google Cloud, Claude is offered as a partner model — a fully managed, serverless model-as-a-service. There is no infrastructure to provision, but there is a gate: each Claude model must be discovered and enabled through Model Garden, Google's catalog of first-party and partner models, before any request will succeed. Your requests then go to Vertex AI endpoints, not to Anthropic. This article walks through the enablement flow and the permission and policy snags that most often stop it.

Prerequisites

Before you open Model Garden, three things must already be true in the project you plan to use:

A Google Cloud project with billing enabled. Partner models are pay-as-you-go; a project without a billing account cannot accept the terms.
The Vertex AI API (aiplatform.googleapis.com) enabled on the project. Enabling an API requires the Service Usage Admin role (roles/serviceusage.serviceUsageAdmin).
The right IAM roles for the person doing the enabling. Google documents two: the Consumer Procurement Entitlement Manager role (roles/consumerprocurement.entitlementManager) to enable partner models in Model Garden, and the Vertex AI User role (roles/aiplatform.user) — which carries the aiplatform.endpoints.predict permission — to actually send prompts afterwards.

Finding and enabling the model

In the Google Cloud console, open Model Garden and search for the Claude model you want — for example Claude Opus 4.8 or Claude Sonnet 5. Each model has its own model card, and each card has an Enable button. Enablement is per model: turning on Opus 4.8 does not turn on Sonnet 5, so repeat the step for every model your teams will use.

Clicking Enable presents Anthropic's Terms of Service. This step has a gotcha that surprises resellers' customers: Anthropic prohibits certain resellers from reselling its products, and Google states plainly that if your Cloud billing account is managed by a prohibited reseller, you will be unable to accept the terms or enable Claude models. If the accept button fails for no apparent reason, check who owns your billing relationship before filing a support ticket.

Org policy gotchas

Two organization-level policies can silently block enablement in locked-down enterprises:

Enabling partner models requires your org policy to allow the Cloud Commerce Consumer Procurement API (cloudcommerceconsumerprocurement.googleapis.com) — this is the machinery behind the terms-acceptance flow.
If the project sits inside an Assured Workloads boundary (Google's controlled-environment product for regulated workloads), enabling Claude via the Marketplace may require violation exceptions for cloudcommerceconsumerprocurement.googleapis.com and commerceagreement.googleapis.com.

For public-sector readers: Google documents that Claude access on this platform meets FedRAMP High requirements and operates within the Google Cloud FedRAMP High authorization boundary. As always, compliance is inherited from your cloud provider — confirm specifics with Google for your case.

Confirm access before writing real code

Once the model card shows as enabled, verify end to end with the smallest possible call. Install the SDK (pip install -U google-cloud-aiplatform "anthropic[vertex]"), authenticate with gcloud auth application-default login, then run:

from anthropic import AnthropicVertex

client = AnthropicVertex(project_id="your-project-id", region="global")
message = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=64,
    messages=[{"role": "user", "content": "Say 'access confirmed'."}],
)
print(message.id, message.content[0].text)

A successful response has a message ID beginning with msg_vrtx_ — your proof the request went through Vertex. A 403 at this point usually means the caller lacks roles/aiplatform.user, or the model was enabled in a different project than the one your credentials point at. Note that on Vertex the model ID is a bare Anthropic-style name (claude-opus-4-8) — no anthropic. prefix, which is a Bedrock convention.

Rule of thumb: enable every model tier you might route to (for example Opus for hard tasks, Haiku for cheap ones) during this setup session. Terms acceptance is the slow, approval-laden step; doing it once beats rediscovering it mid-sprint.

Where to go next

With access confirmed, set up production credentials in service account authentication, run through the project preflight checklist, and make your first real call with the Vertex quickstart.

Sources

← All articles