Claude Models on Google Cloud: Anthropic managed or Vertex?

Claude Models on Google Cloud (Anthropic managed) is the Claude API run by Anthropic but reached through your Google Cloud project, with IAM for access and Marketplace for billing. It gives Google Cloud teams the full Claude API surface without API keys. Moving needs a new Anthropic organisation, new IAM grants and a new SDK client.

What was released

Claude Models on Google Cloud (Anthropic managed) was announced by Anthropic on 22 October 2026 in the Claude API release notes. In Anthropic's words it is "the Claude API operated by Anthropic and reached through your Google Cloud project". Google Cloud IAM controls access, Google Cloud Marketplace handles billing, and fast mode is available for Claude Opus 5.5 on the new route.

It sits alongside the existing offering, now called Claude Models on Google Cloud (Google managed), where Google runs Claude as part of its Agent Platform and your code uses the AnthropicVertex client. Anthropic's comparison page sets the two side by side.

What actually changed

The new route serves the Claude API itself, not Google's prediction endpoints. According to the setup and reference guide:

• Endpoint: requests go to claude.googleapis.com, under a base URL that names your project, the global location and your workspace ID, followed by the normal Claude API path such as /v1/messages.
• Request format: the same as the Claude API. The model goes in the request body, the anthropic-version and anthropic-beta headers stay, and streaming uses "stream": true rather than a separate :streamRawPredict endpoint.
• Authentication: a Google OAuth access token in an Authorization: Bearer header, from Application Default Credentials. There are no API keys, and the new SDK client ignores ANTHROPIC_API_KEY and ANTHROPIC_BASE_URL.
• Access control: three predefined IAM roles on the project - roles/claude.viewer, roles/claude.user and roles/claude.admin. Only the admin role can open the Claude Console.
• Organisation: setup creates a new Anthropic organisation for the project, separate from any you already have, including Claude Enterprise.
• Features: tool use (web search, web fetch, code execution, memory, the MCP connector), Agent Skills, the Files API, prompt caching, structured outputs, the Message Batches API and Claude Managed Agents all work. MCP tunnels and API keys do not.
• Billing and limits: Anthropic meters usage and reports it to your Google Cloud bill in Claude Consumption Units, and it counts toward a Google Cloud spend commitment like other Marketplace purchases. Rate limits are set by Anthropic, not Google Cloud quotas.
• SDKs: a new AnthropicGoogleCloud client (in beta) for Python, TypeScript, C#, Go, Java, PHP and Ruby, plus support in the ant CLI.

What it means for teams building AI products

One API surface across providers. On the Google managed route, the request body, URL and streaming differ from the Claude API, so teams that also call api.anthropic.com keep two code paths. The new route uses the same body, headers and responses, so you change only the client constructor. That matters most for agent code that relies on newer API features such as Managed Agents, Skills or the Files API.

Access becomes an IAM question. There are no keys to rotate or leak, and Cloud Audit Logs records activity under claude.googleapis.com. Two details are easy to miss. Data Access logs, which cover messages, batches and files, are off by default, so turn them on if you need a record of requests (see how to audit what an AI agent did). And roles apply to every workspace in the project, so teams that need separate access need separate projects, each with its own organisation.

Revoking access does not stop running agents. Anthropic's guide states that Managed Agents sessions and scheduled deployments check access only when they are created. Removing a user or rotating a service account leaves their sessions running until you interrupt them and their deployments running until you pause or archive them. Add that review to your offboarding checklist, especially if you use dynamic workflows in Managed Agents.

Nothing carries over. An existing Anthropic organisation can't be converted. API keys, workspaces, console settings, files, skills, batches and Managed Agents resources stay where they are, and a Zero Data Retention arrangement on another organisation does not transfer. Existing discounts can't be applied to past usage, so speak to your account representative before you subscribe.

Data processing changes hands. On this route Anthropic is the data processor, and some supporting services, such as trust and safety systems, process data outside Google Cloud. If your approval to use Claude rested on Google being the processor, that review has to be repeated.

How to move from the Vertex client

The guide sets out a cutover rather than a switch, and both routes can serve traffic in parallel:

1. Subscribe to the Marketplace listing, then enable claude.googleapis.com and anthropic-managed-claude.cloudpartnerservices.goog on every project that calls it.
2. Activate the Anthropic organisation, open the Claude Console and copy the workspace ID (it starts with wrkspc_).
3. Grant roles/claude.user to each calling identity and set ANTHROPIC_GOOGLE_CLOUD_PROJECT and ANTHROPIC_GOOGLE_CLOUD_WORKSPACE_ID.
4. Replace AnthropicVertex with AnthropicGoogleCloud. For models older than Claude 4.6, change the @ before the snapshot date in the model ID to a hyphen.
5. If you use a regional or multi-region endpoint today, note that this route uses global; for US-only inference set inference_geo: "us" on each request.
6. Check the rate limits for the new organisation before moving production traffic, and keep the old route as a fallback, as you would for any provider outage plan.

When not to switch

CodeDTX's view: stay on the Google managed route if your data residency approval depends on regional endpoints and Google as processor, if you rely on Google Cloud quotas you have already negotiated, or if your workload uses only the Messages API and works well today. Wait if you depend on MCP tunnels, on the pre-built self-hosted sandbox workers, or on memory stores in self-hosted environments, none of which are supported yet. Move when you want one Claude API code path across clouds, need Managed Agents or Skills inside Google Cloud, or want Claude spend counted against your Google Cloud commitment through Marketplace.

Frequently asked questions

Is Claude Models on Google Cloud (Anthropic managed) the same as Claude on Vertex AI?

No. The Google managed offering runs Claude on Google's Agent Platform, with the model in the URL and its own request format, and your code uses the AnthropicVertex client. The Anthropic managed offering is the Claude API itself, operated by Anthropic and reached through claude.googleapis.com. Both bill to Google Cloud and use Google credentials, but they need different IAM roles, SDK clients and endpoints.

Can I keep using API keys with the Anthropic managed offering?

No. Anthropic's guide says authentication uses Google Cloud credentials only, and the Claude Console can't create API keys for the organisation that setup creates. Calls carry a short-lived Google OAuth access token from Application Default Credentials, a service account, or Workload Identity Federation. The new SDK client ignores ANTHROPIC_API_KEY, so code that relies on a key will not use it silently.

Will my existing Anthropic organisation and files move across?

No. Setup always creates a new Anthropic organisation for the Google Cloud project, and an existing one can't be converted. Workspaces, console settings, files, skills, batches and Managed Agents resources stay in the old organisation, and a Zero Data Retention arrangement does not carry over. Plan it as a cutover: run both in parallel, recreate what you need, and move workloads gradually.

Does revoking IAM access stop a running Claude agent?

Not on its own. For Claude Managed Agents, access is checked when a session or scheduled deployment is created, not while it runs. A running session continues until you send it a user.interrupt event, and a scheduled deployment continues until you pause, update or archive it. When someone leaves or a service account is rotated, review the sessions and deployments that identity created.

Are all Claude API features available on the Anthropic managed route?

Most are, including tool use, the MCP connector, Agent Skills, the Files API, prompt caching, structured outputs, batches and Claude Managed Agents. Anthropic lists a few exceptions: API keys, MCP tunnels, memory stores on self-hosted environments, and the pre-built self-hosted sandbox workers. Betas limited to selected organisations need the same access granted to the new organisation, so check each feature your code depends on before moving.