Claude on Vertex AI (docs.claude.com)

Anthropic’s Claude models are now accessible through Google Vertex AI with a few platform-specific conventions: the model is specified in the Vertex endpoint URL (not in the request body), and you must include anthropic_version in the request body set to "vertex-2023-10-16". Developers should install Anthropic’s client SDK for their language and authenticate with Google Cloud (gcloud auth application-default login) before making requests. Model availability varies by region—search the Vertex AI Model Garden or “Use Claude 3” docs for the latest list and API model IDs. Operationally this integration matters because it lets teams run Claude models within Google’s infrastructure and compliance boundaries. Vertex offers request-response activity logging (Anthropic recommends keeping a 30-day rolling log for monitoring/misuse investigations; enabling logging doesn’t grant Google or Anthropic access to your content). Starting with Claude Sonnet 4.5, Vertex provides two endpoint types: global (recommended — dynamic routing, max availability, pay-as-you-go only, no premium) and regional (fixed data residency, supports provisioned throughput, 10% pricing premium). Older models (Sonnet 4, Opus 4, etc.) remain on their existing pricing and routing. To use global endpoints, initialize the client with region="global" (or choose a specific region like "us-east1"/"europe-west1" for regional routing).