Claude on Google Cloud: Anthropic-Managed vs Google-Managed, Which to Choose

Claude now comes to Google Cloud customers in two distinct forms. Google Cloud has announced general availability of an Anthropic-managed deployment option that sits beside its existing Google-managed option. Both serve the same frontier models at the same pricing and bill through Google Cloud. The real differences are who acts as data processor, which Claude features you get, and how quickly new ones arrive.

If you already run Claude through Google's Agent Platform (formerly Vertex AI), nothing is forced on you. If you wanted features such as Claude Managed Agents but could not leave your Google Cloud account, this is the option that closes that gap. This post compares the two and says who should choose which.

The two options at a glance

Source for the table: Google Cloud's announcement. Google lists Claude Fable 5.1, Sonnet 5.5, Haiku 5.5 and Opus 5.5 as available in both.

What the Anthropic-managed option gives you

Per Google, the Anthropic-managed deployment provides day-zero access to Claude models and Claude Platform features, including Claude Managed Agents (beta), skills and code execution. Anthropic operates the stack as data processor on Google Cloud infrastructure, while you keep Google Cloud IAM for authentication, Cloud Audit Logging for audit trails, and one Google Cloud invoice.

The commercial hook is commitment drawdown. Many enterprises have negotiated annual Google Cloud commitments, and spend through the Marketplace listing counts against eligible commitments. Procurement teams that could not easily add a second vendor can buy Claude Platform capacity through an account they already have.

Anthropic's documentation shows what changes for developers. The Anthropic-managed option uses a dedicated endpoint host, claude.googleapis.com, with project and workspace identifiers in the path, and the SDKs expose it through AnthropicGoogleCloud clients (for example AnthropicGoogleCloud() in Python). Calls look like the normal Messages API, with anthropic-version: 2023-06-01, rather than Vertex-style rawPredict calls. The ant CLI works against it.

What the Google-managed option keeps

Google positions its own option for organizations with strict governance, compliance and sovereignty needs. Google Cloud is the data processor under existing customer agreements, and you get VPC Service Controls, customer-managed encryption keys and data residency in the US, EU and APAC. It also has native integrations with BigQuery and Model Armor. New Claude Platform features arrive as Google integrates them rather than on day zero.

Anthropic's own docs for the Google-managed path list its gaps plainly. It does not support:

• server-side tools: code execution, web fetch and advisor
• agent infrastructure: Agent Skills, the MCP connector and programmatic tool calling
• Claude Managed Agents
• the Files API and URL sources for images and documents
• Message Batches, Models, Admin, Compliance and Usage and Cost APIs
• the server-side fallbacks parameter

It does support the Messages API, prompt caching, thinking, tool use including bash, browser, computer use, memory, text editor and web search tools, citations and structured outputs, with a 1M-token context window on the current Claude models. Regional and multi-region endpoints cost 10 percent more than global ones, and regional endpoints serve Sonnet 4.6 and earlier only.

If your project depends on features in that unsupported list, that list is your answer.

Which one should you choose?

The legal difference deserves a careful read. Under the Anthropic-managed option, Anthropic is the data processor under Anthropic's Commercial Terms, even though it runs on Google Cloud infrastructure. Regulated teams should have counsel read those terms before moving workloads, because the Google Cloud contract no longer governs the processing.

What this means for builders and teams

For a team already on Google Cloud, the practical effect is that Claude's newest platform capabilities are no longer reserved for the Claude API and AWS-style integrations. We covered the agent side in Claude Managed Agents and dynamic workflows with up to 1,000 subagents, and those features are exactly what the new option brings into a Google Cloud account.

The model lineup matches recent launches: Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 with Mythos 5.1. Since pricing is the same across both options, the decision is not about cost per token. It is about features, governance and legal terms, which is an unusual and healthy kind of cloud choice.

How it fits the wider Google and Anthropic relationship

Google already surfaces Claude inside its own products. In Google's Gemini Agent for Work launch, Claude models are included, and Anthropic is expanding where it sells through its Claude marketplace. A second Google Cloud channel, run by Anthropic itself, means Anthropic keeps the customer relationship and the data-processing role for a growing slice of enterprise usage, while Google keeps the infrastructure revenue and the billing relationship. For buyers, it means more bargaining power, since the same models are available through several contracts.

Migration notes if you are switching

Anthropic's docs describe moving from the Anthropic-managed client to the Google-managed one: replace the AnthropicGoogleCloud import with AnthropicVertex, change the endpoint and request shape to Agent Platform's :rawPredict, pass anthropic_version in the body, use Agent Platform model IDs, and check your code against the unsupported-feature list. Moving the other way means replacing the Vertex client with the Anthropic-managed one and gaining the missing features. In either direction, run a test request before moving production traffic and check quotas, which Google manages for the Google-managed option.

A practical approach is to keep model IDs and prompts in configuration, wrap the client creation in one function, and write a small contract test that exercises every feature you rely on. That makes the choice reversible if governance requirements change.

Questions security and procurement teams will ask

Does my data leave Google Cloud? Google says Anthropic operates the stack on Google Cloud infrastructure. What changes is the contractual processor, not the cloud the servers sit in. Ask Anthropic for the data-flow description and retention terms in writing.

Will my audit trail still work? Google lists Cloud Audit Logging as part of the Anthropic-managed option, so API activity can land in the same logging pipeline as the rest of your Google Cloud estate. Anthropic's docs for the Google-managed path also recommend keeping at least 30 days of request and response logs to investigate misuse, a sensible habit on either option.

Can I run both? Nothing in the announcement prevents it. A common pattern would be Google-managed for regulated, residency-bound workloads and Anthropic-managed for agent and tooling experiments, with a thin internal wrapper deciding which client to call.

What about spend visibility? Both consolidate under Google Cloud billing, so FinOps tooling you already use should see the charges. Check how Marketplace line items are labeled before you build chargeback reports.

What is the risk of waiting? Because models arrive on day zero in both options, waiting costs you only the platform features, not model access.

A short decision checklist

Before choosing, answer four questions in order. First, does any feature you need appear on the Google-managed unsupported list, such as Claude Managed Agents, skills, code execution or the Files API? If yes, the Anthropic-managed option is the practical choice. Second, do your compliance rules require VPC Service Controls, customer-managed encryption keys or Google as the contractual processor? If yes, stay Google-managed. Third, has counsel reviewed Anthropic's Commercial Terms for the processing role? Fourth, have you confirmed how Marketplace spend will draw down your commitment? Writing the answers down turns a vague platform debate into a decision you can defend later.

Limits of what we know

The announcement does not give a price change, region list for the Anthropic-managed option, or compliance certification list. It says the pricing is standard and eligible commitments are drawn down, but "eligible" is defined by your Google Cloud agreement. Regional availability and certifications for the Anthropic-managed stack should be confirmed with Google or Anthropic before you commit regulated workloads. The Google announcement page we read was dated October 23, 2026 on the page itself; we cite it as the announcement for this launch.

Related reading

• Claude Managed Agents dynamic workflows
• Claude Opus 5.5 launch
• Claude Sonnet 5.5 building guide
• Claude Haiku 5.5 launch
• Claude Fable 5.1 and Mythos 5.1
• Google Gemini Agent for Work with Claude models
• Anthropic Claude marketplace expansion
• Official: Google Cloud announcement and Claude docs for Google-managed deployment

Details are from Google Cloud's announcement and Anthropic's documentation as of October 22, 2026 and may change.