Vertex AI vs. the direct Anthropic API
Calling Claude through Google Cloud Vertex AI differs from Anthropic's own client SDKs in the same broad shape that Bedrock differs -- a cloud-native hosting layer wrapping the identical underlying model:
Authentication
The direct Anthropic API uses a single portable ANTHROPIC_API_KEY bearer header. Vertex AI uses standard Google Cloud IAM: a service account key, Application Default Credentials (ADC) resolved from the environment, or a short-lived OAuth2 access token obtained via gcloud auth print-access-token or a workload identity binding. There is no portable API key equivalent -- credentials are scoped to a GCP project and IAM role.
Model addressing
Vertex AI addresses Claude models as publisher models: publishers/anthropic/models/<model-id>, resolved within a specific GCP project and region. This is a third ID convention, distinct from both the direct API's short public model names and Bedrock's anthropic.<model-id>-v1:0 ARN-style IDs.
Billing
Usage is billed through the customer's Google Cloud billing account, consolidated with the rest of their GCP spend, rather than through a separate Anthropic Console billing relationship -- the same cloud-native-billing pattern Bedrock offers via AWS billing.
Regions
Claude models on Vertex AI are available in a defined, model-specific subset of GCP regions rather than being globally available by default -- check Google's current per-model regional availability table before deploying.
Source: cloud.google.com/vertex-ai/generative-ai/docs/partner-models/use-claude. See also the full fact catalog for every cited claim, and subagentbedrock.com's Bedrock-vs-API page for the AWS-side parallel.