Claude Can Now Be Used as the "Anthropic-Operated Claude API" from a Google Cloud Project - IAM for Access, Marketplace for Billing

On October 22, 2026, Anthropic launched Claude Models on Google Cloud (Anthropic managed), which lets customers reach the Claude API run by Anthropic itself through a Google Cloud project. Access is controlled by Google Cloud IAM, billing goes through Google Cloud Marketplace, and fast mode for Opus 5.5 is available. We cover how it differs from the existing Google-operated option and what to watch for when moving.

Anthropic Claude Google-Cloud API enterprise cloud

In its October 22, 2026 release notes, Anthropic announced the launch of a new offering, “Claude Models on Google Cloud (Anthropic managed).” Customers connect to the Claude API that Anthropic runs, going through their own Google Cloud project; Google Cloud IAM handles access control and Google Cloud Marketplace handles billing ¹. Google Cloud also announced general availability of this option on its official blog ⁴.

Google Cloud also offers Claude in a form where Google runs the models as part of its own Agent Platform. Anthropic’s documentation now presents the two side by side as “Anthropic managed” and “Google managed” offerings ³. Even though both put Claude charges on the same Google Cloud bill, they differ in who operates the service, in the shape of the API, and in which terms apply.

How the two offerings differ

Drawing on Anthropic’s comparison table ³ and Google Cloud’s announcement ⁴, the main differences are as follows.

Aspect	Anthropic managed (new)	Google managed (existing)
Operator	Anthropic (the Claude API itself)	Google Cloud (part of Agent Platform)
API shape	Identical to the Claude API (`/v1/...`, `model` in the request body)	Agent Platform prediction endpoints (`model` in the URL, streaming on a separate endpoint)
Data processor / governing terms	Anthropic, under its Commercial Terms	Google Cloud, per the customer’s current Google Cloud agreement
Anthropic organization	A new organization is created per project	None
Who manages rate limits	Anthropic	Google Cloud
Location	`global`	Global, multi-region, or regional
Billing	Google Cloud Marketplace	Charged by Google Cloud as usage of Agent Platform

There are shared points too. Both use Google Cloud credentials, Application Default Credentials included, and charges for both appear on the Google Cloud bill. Anthropic’s SDKs support both ³. Google Cloud says that under either option the models are available on day zero at the same price, and it lists Claude Fable 5.1, Sonnet 5.5, Haiku 5.5, and Opus 5.5 among the available models ⁴.

According to Google Cloud, the choice comes down to two things: which company processes the data, and whether new Claude Platform features are needed on the day they ship ⁴. With Anthropic managed, capabilities like Claude Managed Agents (in beta), code execution, and skills arrive at launch. With Google managed, new features are added gradually as Google integrates them into its own infrastructure. For Google managed, Google Cloud highlights VPC Service Controls, data residency in the US, the EU, and Asia-Pacific, and integration with BigQuery and Model Armor.

Fast mode and Managed Agents are available

Anthropic’s documentation for the Anthropic managed offering lists features that work the same way they do on Anthropic’s own API ². They include Opus 5.5’s fast mode; tools such as code execution, web fetch, web search, the MCP connector, and the memory tool; the Files API and Agent Skills; prompt caching; extended thinking; structured outputs; and the Message Batches API. It also says context windows are the same size as on the direct API. The documentation notes, however, that the list is not exhaustive and does not indicate whether features left off it will work.

The October 22 release notes explicitly state that Opus 5.5’s fast mode can be used with the Anthropic managed offering ¹.

Some things are not supported. The biggest is API keys: authentication works only with Google Cloud credentials. MCP tunnels, memory stores in self-hosted environments, and the ready-made self-hosted workers bundled with the SDKs and CLI are also excluded ².

Claude Managed Agents comes with a permissions caveat. Permissions are verified only at the moment a session or scheduled deployment is set up, and not again during its run. Revoking the Google Cloud access of the person who created a session therefore does not stop that session ². The documentation asks administrators who offboard a user or rotate service account credentials to go through whatever that identity started, both sessions and deployments.

Permissions, billing, and data handling

Permissions come from three IAM roles (roles/claude.viewer, roles/claude.user, and roles/claude.admin) granted on the project. The Claude Console cannot be reached by signing in with an email address; it is opened from the Google Cloud console, and only holders of roles/claude.admin can open it ². Because a role granted on a project covers all of that project’s workspaces, the documentation recommends a separate Google Cloud project per team where teams need distinct access. The SDK clients for this offering are in beta.

Billing is handled via Google Cloud Marketplace, while Anthropic meters consumption and passes it to Google in units called Claude Consumption Units, and the charges show up on the Google Cloud bill, with no separate invoice from Anthropic. Existing Google Cloud spend commitments are drawn down by this usage just as they are by other Marketplace purchases ². Specific prices are not given: Google Cloud’s comparison table only says “standard pricing” ⁴, and Anthropic’s documentation does not list prices either.

On data, the two companies emphasize different points. Google Cloud says Anthropic runs the service, acting as data processor, on top of Google’s infrastructure ⁴. Anthropic’s documentation, besides naming Anthropic as data processor, states that certain supporting systems, its trust and safety systems among them, handle data outside Google Cloud ². Data retention follows the same policy as the direct API, and customers can ask for Zero Data Retention (ZDR). A ZDR arrangement made for another Anthropic organization does not carry over.

On where inference runs, the documentation says to keep the location at global and, when US-only inference is needed, to set inference_geo: "us" on each request ². The US is the only example given, and the documentation does not say whether other regions, including Japan, can be specified.

Existing Anthropic organizations have to switch over

An organization that already calls api.anthropic.com with API keys cannot simply move into this offering. Existing Anthropic organizations cannot be converted; setup creates a new organization for the Google Cloud project, separate even from a Claude Enterprise organization ².

The move therefore has to be planned as a switch to a fresh organization. Nothing carries over in terms of API keys, Console settings, or workspaces, and resources such as files, batches, skills, and Managed Agents objects remain in the original organization. Since each organization stands on its own, both can serve traffic at the same time and workloads can be moved in stages. The documentation also says that customers with an existing discount should contact their Anthropic or Google Cloud representative before setup, as discounts cannot be applied retroactively ².

Moving from Google managed means replacing the SDK client AnthropicVertex with AnthropicGoogleCloud. Since requests take the same form as the Claude API, the URL, headers, and body change as well ².

Which one to choose

On October 21, it was announced that OpenAI’s models will also be offered in Google Cloud’s model garden. Options for using models from several vendors within a Google Cloud contract and IAM setup are expanding.

For Claude specifically, the deciding factors are fairly clear. If you want new Claude Platform features and Managed Agents as soon as they ship, or want the same request format as Anthropic’s direct API, Anthropic managed is the candidate. If you want Google Cloud to be the data processor under the agreement you already have with it, or need region-specific data residency or VPC Service Controls, the existing Google managed offering is likely the better fit. Either way, working out in advance what to do about existing API-key-based operations, and about files and Managed Agents resources left in the original organization, makes the switch easier to plan.