Advanced

Google Vertex AI (Claude)

Claude is available through Google Cloud's Vertex AI platform as part of Vertex's Model Garden. Like AWS Bedrock, this lets GCP-native organisations use Claude without a separate Anthropic account — billing, IAM, VPC, and compliance controls all flow through GCP.

Why Use Vertex Instead of Direct API?

Reasons to use Vertex

GCP billing consolidation: Charge Claude usage to your GCP project — apply CUD commitments and billing controls
Service account auth: Use Google service accounts and ADC — no separate API key management
VPC-SC: Route Claude calls through VPC Service Controls — data stays within your GCP perimeter
Data residency: Select a GCP region — data stays within that region
GCP IAM: Grant Claude access via IAM roles consistent with your other GCP services
Integrated with Vertex pipelines: Use Claude as a step in ML Pipelines, Feature Store workflows, or BigQuery ML

Reasons to use direct API

Latest Claude models available immediately (Vertex lags on new releases)
Simpler setup — API key vs service account + project configuration
All Anthropic API features available (some may lag on Vertex)
Cloud-agnostic code — not locked to GCP

Enabling Claude on Vertex AI

To use Claude through Vertex AI:

Open Google Cloud Console → Vertex AI → Model Garden
Search for "Claude" — you'll see available Claude models (Haiku, Sonnet, Opus variants)
Click "Enable" on the Claude model card — this enables the API for your project
Accept the Anthropic terms of service (shown during enablement)
Enable the Vertex AI API for your project if not already enabled: gcloud services enable aiplatform.googleapis.com

Once enabled, Claude is available via the Vertex AI Prediction API endpoint for your project and region.

Authentication: Service Accounts and ADC

Vertex AI uses Google authentication, not API keys. There are two approaches:

Application Default Credentials (ADC): The recommended approach for code running on GCP infrastructure (GCE, GKE, Cloud Run, Cloud Functions). The runtime environment automatically provides credentials. Run gcloud auth application-default login for local development.
Service account key file: For workloads running outside GCP. Create a service account with the Vertex AI User role, download a JSON key, and set the GOOGLE_APPLICATION_CREDENTIALS environment variable to the key path. Avoid this in production if possible — ADC is more secure.

Vertex SDK vs Anthropic SDK

You have two options for calling Claude on Vertex:

Anthropic SDK with Vertex backend

Anthropic's Python and TypeScript SDKs support a Vertex AI backend. Set the project, region, and use AnthropicVertex instead of Anthropic. The request format is identical to the direct API.

Best for: teams already using the Anthropic SDK who want minimal code changes when switching to Vertex.

Google Cloud Vertex AI SDK

Use google-cloud-aiplatform library with the generative_models module. Consistent with other Vertex AI model calls; supports GCP-native features like Vertex Pipelines integration.

Best for: teams building GCP-native workflows where consistency with other Vertex models matters.

Integration with GCP Services

Running Claude on Vertex unlocks native integrations with other GCP services:

BigQuery: Use Claude as a SQL generation assistant or data analysis tool within BigQuery workflows — trigger Claude from BigQuery ML or via Cloud Functions
Cloud Storage: Reference documents in GCS buckets directly in Vertex pipeline steps — Claude can process files pulled from GCS without manual download
Vertex Pipelines (Kubeflow): Include Claude as a pipeline component — useful for ML workflows that involve text processing, labelling, or evaluation steps
Cloud Logging / Cloud Monitoring: All Vertex AI API calls log to Cloud Logging automatically — queryable for auditing, cost analysis, and debugging

Cost and Quota

Claude on Vertex is priced per token (input + output), similar to Bedrock and direct API — check the Vertex AI pricing page for current rates
Quotas are set at the project level — request quota increases via the GCP Console if you hit limits
Use GCP budget alerts and cost allocation labels to monitor Claude spend alongside other GCP costs
Vertex charges are separate from any Anthropic direct API spend — track each independently if using both

Checklist: Do You Understand This?

Vertex AI hosts Claude within GCP — enables GCP billing, service account auth, VPC-SC, and data residency
Enable Claude via Model Garden in the GCP Console; accept Anthropic ToS during enablement
Use ADC for authentication — avoid service account key files where possible
Anthropic SDK supports a Vertex backend (AnthropicVertex) with the same request format as direct API
New Claude versions lag on Vertex vs direct API — check Model Garden for current available versions
Native integrations with BigQuery, Cloud Storage, and Vertex Pipelines make it useful for GCP-native ML workflows