Claude on Google Cloud: Vertex AI or Anthropic managed?

Claude runs from a Google Cloud project in 2 ways. Google managed is Claude in Vertex AI, which Google now calls Agent Platform: Google runs the model, sets your quota and invoices it like any other Agent Platform call. Anthropic managed, launched on 22 October 2026, is the Claude API itself, run by Anthropic, reached at claude.googleapis.com with Google IAM for access and Google Cloud Marketplace for the bill. The token prices are the same. The features are not: the Anthropic managed route has the Batch and Files APIs, code execution, Agent Skills and Claude Managed Agents, and the Vertex AI route has none of them. Vertex AI has regional endpoints and provisioned throughput; on the Anthropic managed route the location is always global.

This is the same split AWS has, where Amazon Bedrock sits beside Claude Platform on AWS; our Bedrock vs the direct API piece covers that side.

Which one are you on now?

Most teams that "use Claude on GCP" set it up before October 2026, so they are on Vertex AI. Anthropic's overview page lists the values that settle it, and any one of them is enough:

If your code calls api.anthropic.com with an API key, you are on neither. That is the plain Claude API, billed by Anthropic.

What each route can call

This is where the choice is usually made. Anthropic lists, for each route, what works and what does not:

If your agent needs a tool that runs on Anthropic's side, such as code execution, a Skill, or the MCP connector that lets Claude call a remote MCP server for you, Vertex AI cannot do it, and you would have to run that piece yourself. The same goes for batches: there is no Message Batches API on Vertex AI, only Google's own batch prediction.

The lists can move. The Anthropic managed route is the Claude API, but not every Claude API feature is there (MCP tunnels are not, for one), and Anthropic's list does not say whether an unlisted feature works, and a beta limited to selected organizations needs the same access granted to the new organization. On Vertex AI, Google also sets the retirement dates for the models it serves, and they can differ from Anthropic's schedule.

Same price, different invoice

On the global endpoint Google's price page and Anthropic's agree to the cent. For Claude Opus 5.5, per million tokens:

The batch row is the one place they part. Anthropic's Batch API halves the price, to $2 and $10. Google lists $2.50 and $12.50 for Opus 5.5 batch work on Agent Platform. If you run large overnight jobs, that gap is worth checking against your own volumes; the Opus 5.5 price page has the current figures and the cost calculator does the sum.

Keeping data in one place costs about the same on both. On Vertex AI the US and EU multi-region endpoints, and single regions, cost 10% more than global for Claude Sonnet 4.5 and later. On the Anthropic managed route the location stays global, and for Claude 4.6 and later models you ask for US-only inference by setting inference_geo: "us" on a request, which multiplies every token price by 1.1. Say a month of Opus 5.5 work is 200 million tokens in and 20 million out: $1,200 on global routing, $1,320 kept in the US, by either route.

The invoices look different. Vertex AI usage shows up as Agent Platform usage. The Anthropic managed route shows up as a Marketplace line in Claude Consumption Units, at $0.01 each, with no separate bill from Anthropic, and it counts toward a Google Cloud spend commitment the way other Marketplace purchases do. That commitment can be a reason to pick this route over the plain Claude API: same API, paid from money already promised to Google.

Rate limits and where your data goes

The two routes are run by different companies, and that shows up in 3 places:

• Rate limits. Google sets Vertex AI quotas, and you request increases as Google Cloud's quota page for Claude models describes. Anthropic sets limits on the Anthropic managed route, and you raise them through Anthropic sales. A team that hits Vertex AI quota at peak hours has a second, separate pool to draw from.
• Who processes the data. On Vertex AI, Google Cloud's data rules apply. On the Anthropic managed route Anthropic is the data processor, with the Claude API's retention policy; some supporting services, such as trust and safety systems, run outside Google Cloud. Zero data retention is on request, and an arrangement on another Anthropic organization does not carry over.
• Logs. Vertex AI has request-response logging for prompts and answers. The Anthropic managed route writes to Cloud Audit Logs under claude.googleapis.com; Data Access logs, which cover messages, are off until you turn them on.

For a hard rule such as "requests never leave europe-west1", only Vertex AI can do it, and only for Claude Sonnet 4.6 and earlier: Google's regional endpoints do not serve the newer models. Claude Opus 5.5 and Sonnet 5.5 run on global or a US or EU multi-region endpoint.

The code: a client name and a URL

Anthropic's SDKs cover both routes (the Anthropic managed client is in beta), so a move is small. Here is the same Python request on each:

# Vertex AI (Google managed): pip install -U "anthropic[vertex]"
from anthropic import AnthropicVertex

client = AnthropicVertex(project_id="my-project", region="global")  # or "us", "eu", "us-east5"

# Anthropic managed: pip install -U "anthropic[google_cloud]"
from anthropic import AnthropicGoogleCloud

client = AnthropicGoogleCloud()  # reads ANTHROPIC_GOOGLE_CLOUD_PROJECT and ANTHROPIC_GOOGLE_CLOUD_WORKSPACE_ID

# Either way, the call is the same
message = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Summarise yesterday's failed jobs"}],
)
print(message.content[0].text)

Underneath, the requests differ more. Vertex AI puts the model in the URL and anthropic_version in the body, and streams from a separate :streamRawPredict endpoint. The Anthropic managed route is a normal Claude API request: model in the body, the anthropic-version header, "stream": true, and the workspace ID in the path instead of a header. Model IDs match from Claude 4.6 onward; older ones on Vertex AI swap the last hyphen for @, so claude-haiku-4-5-20251001 becomes claude-haiku-4-5@20251001.

Setting up the Anthropic managed route

It takes more setup than Vertex AI, because it creates an Anthropic organization for your project:

1. Find Claude Models on Google Cloud (Anthropic managed) in Google Cloud Marketplace and subscribe. You need permission to buy on the billing account.
2. Enable 2 services on each project that calls it: claude.googleapis.com, which serves the API and defines its permissions, and anthropic-managed-claude.cloudpartnerservices.goog, which ties usage to your subscription.
3. Click Activate, then fill in the organization details from the emailed link and accept Anthropic's terms.
4. Click Open Claude Console, copy the workspace ID (it starts wrkspc_) and give it to the developers.
5. Grant roles/claude.user to each identity that calls Claude. Only roles/claude.admin can open the Claude Console.

Three details catch people out. The new organization is separate from any Anthropic organization you already have, Claude Enterprise included, so files, Skills, batches and agents stay where they were; Anthropic treats it as a cutover, and the two can run side by side while you move traffic. Roles are granted on the project and apply to every workspace in it, so teams that need separate access need separate projects, each with its own organization. And Managed Agents check access only when a session or scheduled deployment is created: revoking someone's IAM role does not stop a session they started, so review what an identity created when you remove it.

If you have a negotiated discount with Anthropic or Google, talk to your account team before you subscribe. Anthropic says discounts cannot be applied to past usage.

Which one to pick

• You call Claude for chat, extraction or classification, and need EU or single-region processing. Stay on Vertex AI. Its multi-region and regional endpoints are the only way to pin the region, and you need none of the missing features.
• You build agents on Anthropic's server-side tools, Skills, the Files API or Managed Agents. Use the Anthropic managed route. Those features exist only there on Google Cloud.
• You run big batch jobs. Compare Anthropic's Batch API on the Anthropic managed route with Google's batch price for your model and volume; the gap was 25% for Opus 5.5 when we checked.
• You already pay Anthropic directly but want the spend on your Google Cloud commitment. Use the Anthropic managed route. It is the same API, so code changes are the client and the credentials.
• You want both. Nothing stops you. The roles are separate, the quotas are separate and the bill is one. You can route region-bound traffic to Vertex AI and agent workloads to the Anthropic managed route.

If you are weighing the same question on Azure, Azure OpenAI vs the OpenAI API walks through how Microsoft's route differs from buying direct.