Claude and OpenAI models through Google Cloud: what changes

On 22 October 2026 Anthropic launched Claude Models on Google Cloud (Anthropic managed): in Anthropic's words, "the Claude API operated by Anthropic and reached through your Google Cloud project", with Google Cloud IAM controlling access and Google Cloud Marketplace handling billing¹. It is not Claude on Vertex AI. That offering, which Anthropic's documentation now calls Claude Models on Google Cloud (Google managed), is run by Google as part of its Agent Platform². A day earlier, OpenAI and Google announced OpenAI's frontier models and Codex for Google Cloud's Model Garden, in private preview, transacted through Google Cloud Marketplace and able to draw down Google Cloud commitments³⁴.

For a developer the choice changes the endpoint, the authentication, the feature set and the quotas. For a CFO it changes whose invoice the spend appears on, which commitment it burns down, and who processes the data. This comparison covers both, and the cases where each option loses. It extends Claude on the Anthropic API, Bedrock or Google Cloud, written before the Anthropic-managed offering existed.

The options side by side

The cells come from Anthropic's pages on the two Google Cloud offerings⁵⁶², its Bedrock and Claude Platform on AWS pages⁷⁸, and the two announcements about OpenAI models⁴³. Anthropic also offers an AWS counterpart of the new offering: Claude Platform on AWS, which Anthropic operates, with AWS handling authentication, IAM access control and billing through AWS Marketplace⁸.

Who runs the model and who sees the data

This is the row a security review asks about first.

Anthropic managed. "Anthropic is the data processor" for this offering, and Anthropic notes that "some supporting services, such as Anthropic's trust and safety systems, process data outside Google Cloud"⁵. Data retention follows the Claude API's policy; Zero Data Retention is available on request, and a ZDR arrangement on another Anthropic organization "doesn't carry over"⁵. Google Cloud's gateway at claude.googleapis.com checks your credentials and IAM permission, then forwards the request to Anthropic⁵.

Google managed. "Google Cloud governs data handling" for Claude on Agent Platform, and prompt and completion logging is Agent Platform's own request-response logging service⁶. If your requirement is that the cloud provider is the only party processing inference, this is the Google Cloud option that fits it, as Bedrock is on AWS: Anthropic describes Bedrock as running "entirely on AWS-controlled infrastructure with AWS as the operating party"⁸.

OpenAI on Google Cloud. Google's announcement says OpenAI models "run on dedicated Google Cloud infrastructure", that queries are routed through isolated gateways within Google Cloud, and that prompts, context and outputs "remain entirely within your designated data perimeter and are never used to train foundation models"⁴. It also lists Zero Data Retention options and Confidential Compute as "coming soon"⁴. Neither announcement says which company operates the serving stack, so ask before you file the processing record.

Features: where the Google-managed and Bedrock versions lose

The Anthropic-managed offering lists, as working the way they do on the Claude API: fast mode for Claude Opus 5.5, tool use including web search, web fetch, code execution, the memory tool, the advisor tool and the MCP connector, Agent Skills and the Files API, prompt caching, extended thinking, server-side fallback, structured outputs, token counting, the Message Batches API and the Models API, plus Claude Managed Agents⁵. What it does not support: API keys, MCP tunnels, memory stores on self-hosted environments, and the pre-built self-hosted sandbox workers in the SDKs and CLI⁵. Context windows match the Claude API⁵.

The Google-managed offering does not support URL sources for images and documents, the Files API, code execution, web fetch, the advisor tool, Agent Skills, the MCP connector, programmatic tool calling, the Message Batches, Models, Admin, Compliance and Usage and Cost APIs, Claude Managed Agents, or server-side fallback⁶. Anthropic's Bedrock page lists a near-identical gap, adding web search, automatic prompt caching and the newest computer-use and browser-use toolsets⁷.

The practical reading: if your code uses batches, files, server tools or Managed Agents, the Google-managed and Bedrock routes need a rewrite or a second path, and the Anthropic-managed route does not.

Quotas, locations and residency pricing

Rate limits. On the Anthropic-managed offering, "Anthropic manages rate limits"; higher limits come from your Anthropic account representative or sales⁵. On the Google-managed offering, Google Cloud manages quotas and quota increases⁶. Check limits before moving production traffic; Anthropic's migration steps say the same⁵.

Location. The Anthropic-managed offering uses the global location, which is separate from where inference runs; for US-only inference you set inference_geo: "us" per request⁵, which carries a 1.1x multiplier on all token prices for Claude 4.6 and later models on the Claude API, Claude Platform on AWS and this offering⁹. The Google-managed offering has global, multi-region (us, eu) and regional endpoints; multi-region and regional endpoints cost 10 percent more than global, and regional endpoints serve Claude Sonnet 4.6 and earlier only⁶. Bedrock's regional endpoints also carry a 10 percent premium⁷.

Lifecycle. Google sets the retirement dates for models on the Google-managed offering, and they can differ from the Claude API's; Anthropic's own dates apply to the Anthropic-operated platforms⁶.

Billing and commitments: the CFO's view

The reason most enterprises will look at these offerings is the invoice, so the details matter:

• Anthropic managed. It "bills through Google Cloud Marketplace". Anthropic meters usage and reports it to Google Cloud in Claude Consumption Units, where one unit is $0.01; charges appear on your Google Cloud bill and "Anthropic doesn't send a separate invoice". Usage and cost are visible in the Claude Console. And: "If you have a Google Cloud spend commitment, usage counts toward it the way other Marketplace purchases do"⁵.
• Google managed. Google Cloud invoices it as Agent Platform usage, at Google Cloud's published prices, which depend on the endpoint type⁶.
• OpenAI on Google Cloud. OpenAI says enterprises with Google Cloud spend commitments "will be able to draw down their GCP commitments for these purchases"³. Google adds that usage counts toward Flexible Savings Plans across model providers⁴.

Two cautions. First, setup of the Anthropic-managed offering creates a new Anthropic organization, separate from any you have; Anthropic asks customers with existing negotiated pricing to contact their Anthropic or Google Cloud representative before starting, because those terms "can't be applied to past usage"⁵. Do that before the first request, not after the first bill. Second, a commitment drawdown is not a lower rate: it moves spend you already promised to Google onto AI usage. Whether that is worth it depends on how much of the commitment you would otherwise leave unused.

Moving a workload: what the code change is

From the direct Claude API, three things change and the request body does not⁵:

• Base URL: https://claude.googleapis.com/v1/projects/{project}/locations/global/workspaces/{workspace}/invoke, then the usual path such as /v1/messages.
• Authentication: an Authorization: Bearer header with a Google access token; no x-api-key.
• Workspace: in the URL, not in an anthropic-workspace-id header.

# pip install -U "anthropic[google_cloud]"
# Env: ANTHROPIC_GOOGLE_CLOUD_PROJECT, ANTHROPIC_GOOGLE_CLOUD_WORKSPACE_ID (starts wrkspc_)
# The caller needs roles/claude.user on the project; credentials come from ADC.
from anthropic import AnthropicGoogleCloud

client = AnthropicGoogleCloud()
message = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello, Claude"}],
)
print(message.content)

The SDK client for this offering is in beta, and it ignores ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN and ANTHROPIC_BASE_URL⁵. Before cut-over: enable claude.googleapis.com and the Marketplace service on every calling project, grant roles/claude.user to each calling identity, recreate workspaces, and remember that files, skills, batches and Managed Agents resources stay in the old organization⁵. Turn on Data Access audit logs for claude.googleapis.com if you want a record of requests; they are off by default⁵.

From Vertex AI, the change is the client class (AnthropicVertex to AnthropicGoogleCloud), the model id format for models before Claude 4.6 (claude-haiku-4-5@20251001 becomes claude-haiku-4-5-20251001), and a check of the endpoint location you rely on today⁵.

How to choose

For OpenAI models outside Google Cloud, the other cloud route is Azure; OpenAI's announcement states that "Microsoft remains OpenAI's primary cloud partner"³. The trade-offs there are in Azure OpenAI vs the OpenAI API, and running several clouds behind one interface is covered in one gateway across Azure, Bedrock and more.

Which of these INFRO connects

INFRO is a control plane that sits between your applications and provider accounts you already hold; the provider or the cloud bills you directly, and INFRO charges a separate subscription. Today INFRO connects Claude through your own Anthropic account and through Google Vertex AI, that is the Google-managed offering: you give INFRO a service-account key for your project, it signs short-lived access tokens with it, and the account needs only roles/aiplatform.user. Your application names anthropic/claude-opus-5.5, and a route can order or weight the two connections, so a failure on one moves the request to the other before the response starts.

What INFRO does not connect yet, stated plainly: the Anthropic-managed Google Cloud offering at claude.googleapis.com, Claude in Amazon Bedrock, and OpenAI models in Model Garden. If you move Claude traffic to the Anthropic-managed offering, calls made there do not pass through INFRO and do not appear in its usage or budgets. INFRO also does not change which commitment your spend draws down: that is decided by the account the provider or cloud bills.