Claude on Vertex AI is serverless — there is nothing to provision — but there is a chain of project-level prerequisites, and a failure anywhere in the chain surfaces as a confusing error at call time. This checklist orders them the way dependencies actually flow. Work top to bottom in the Google Cloud project you intend to use (a dedicated project per environment is a good default, since quotas and audit logs are easier to reason about per project).
The checklist
1. Billing is enabled on the project. Partner models are pay-as-you-go against your billing account. No billing, no terms acceptance, no calls. One reseller-specific trap: Anthropic prohibits certain resellers from reselling its products, so if your billing account is managed by a prohibited reseller, Google states you will be unable to accept the Terms of Service or enable Claude at all. Surface this with procurement before the engineering sprint starts.
2. The Vertex AI API is enabled. The service is aiplatform.googleapis.com. Enabling it requires the Service Usage Admin role (roles/serviceusage.serviceUsageAdmin).
3. Org policy allows the procurement API. Enabling partner models depends on the Cloud Commerce Consumer Procurement API (cloudcommerceconsumerprocurement.googleapis.com). Locked-down organizations frequently restrict which services may be enabled; if yours does, get this one allowlisted. Inside an Assured Workloads boundary, enabling Claude via the Marketplace may additionally require violation exceptions for cloudcommerceconsumerprocurement.googleapis.com and commerceagreement.googleapis.com.
4. Check endpoint-related org policies. If your organization enforces constraints/gcp.restrictEndpointUsage to block the global endpoint, plan for regional endpoints from day one — that changes your pricing (+10%) and your quota buckets. See the endpoint decision guide.
5. IAM roles are bound. The admin who enables models needs roles/consumerprocurement.entitlementManager; anyone or anything that will send prompts needs roles/aiplatform.user (it carries the required aiplatform.endpoints.predict permission). Details and job-function mapping in the IAM roles reference.
6. Each Claude model is enabled in Model Garden. Find the model card, click Enable, accept Anthropic's terms — once per model, per project. The full walkthrough is in the Model Garden setup guide.
7. Quota is sanity-checked. Before load-testing, look at the Quotas & System Limits page in the console for the Claude quotas in your chosen location. Default maximums vary by account, and in some cases access can be restricted. If your launch math exceeds the defaults, file the increase now — see requesting a quota increase.
8. A smoke test passes. Authenticate (gcloud auth application-default login for a human; a service account for automation) and make one minimal call:
from anthropic import AnthropicVertex
client = AnthropicVertex(project_id="your-project-id", region="global")
r = client.messages.create(
model="claude-haiku-4-5@20251001",
max_tokens=32,
messages=[{"role": "user", "content": "ping"}],
)
print(r.id) # msg_vrtx_... means Vertex answeredThe three gotchas that cause the most tickets
One more expectation-setting note for stakeholders: enabling Claude on Vertex gets you the core Messages API surface — streaming, tool use, vision, thinking, prompt caching. Anthropic's Message Batches API and Files API are not part of the Vertex offering (Google provides its own Vertex batch prediction mechanism for Claude instead), so if a team's design assumes those, catch it during preflight, not after. The full picture is in platform feature gaps.
Finally, if compliance or security review is on your critical path, decide now whether Data Access audit logs and a VPC Service Controls perimeter are required for this project — both are far easier to enable before traffic starts flowing than to retrofit, and neither is on by default.
Where to go next
With preflight green, follow the first-API-call guide, then wire production credentials per service account authentication. The site-wide readiness checklist covers the non-GCP parts of going live.