What changed
Anthropic’s blog says Claude Models on Google Cloud (Anthropic managed) is generally available, with Google Cloud IAM credentials, billing, and commitment retirement. The same post says billing runs through a single Google Cloud invoice and that Claude usage draws down an existing Google Cloud commitment. The article says the offering is generally available today. It prints no calendar date.
On 22 October 2026, @claudeai posted at 21:21 UTC and ClaudeDevs posted at 21:27 UTC. Those are the times the posts appeared. ClaudeDevs said a Google Cloud account can now use the full Claude API, including Claude Managed Agents, code execution, and Opus 5.5 fast mode, with new features as they launch. That post says the path uses Google Cloud IAM credentials, usage draws down a Google Cloud commitment, and US and Global endpoints are available. The @claudeai post says the offering is generally available and that usage draws down a Google Cloud commitment.
General availability in the blog and those posts refers to the offering. The setup guide marks the AnthropicGoogleCloud SDK client as beta. That client reads Application Default Credentials and builds the base URL from the project, location, and workspace ID. The Python extra named in the guide is anthropic[google_cloud]. An ant CLI version without support for this offering ignores the ANTHROPIC_GOOGLE_CLOUD_* variables.
Authentication is Google Cloud IAM, granted on the project. The guide’s predefined roles are roles/claude.viewer, roles/claude.user, and roles/claude.admin. A caller needs roles/claude.user or roles/claude.admin. Only roles/claude.admin can open the Claude Console. Viewer is read-only. Every calling project enables claude.googleapis.com, which serves the API and defines the IAM permissions, and anthropic-managed-claude.cloudpartnerservices.goog, which attributes usage to the Marketplace subscription. The organization this setup creates cannot have API keys. The SDK ignores ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, and ANTHROPIC_BASE_URL.
The blog uses the same US and Global endpoints wording. The setup guide tells callers to leave location at global. For US-only inference, set inference_geo to "us" on each request. The guide says location is separate from where inference runs, and it points to a Data residency page for that setting.
The blog says Google Cloud customers now receive the full set of Claude API features, and that new features and betas ship the same day they go live on the native Claude Platform. The @claudeai post says the same about same-day features and betas. It names the Skills API, code execution, and Opus 5.5 fast mode. On the setup guide, the corresponding line is Agent Skills and the Files API. The blog and the setup guide do not use Skills API as a product name of their own.
The guide says that if a beta is limited to selected organizations on the Claude API, the Anthropic organization this setup creates needs the same access. The announcement and the setup guide do not say how that access is granted.
The blog says Claude Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5 are available, with later models as they launch. It says fast mode for Opus 5.5 is generally available on this offering. Anthropic describes that fast mode as up to 2.5x faster output-token generation, at $8 per million input tokens and $40 per million output tokens. The setup guide confirms fast mode for Opus 5.5. It does not state the speed claim or those prices. Neither page states ordinary token prices.
The guide uses Claude API model IDs, with claude-opus-5-5 as its example. A request for a model outside that organization’s inference configuration returns an error. The guide does not say which of the four models a given organization receives.
The setup guide confirms a specific list and says that list does not cover every Claude API feature. It does not say whether an unlisted feature works. The blog marks Claude Managed Agents as beta, advisor strategy as beta, and the MCP connector as beta. The guide describes Claude Managed Agents, the advisor tool, and the MCP connector as available and does not repeat those labels. Managed Agents on this path include memory stores. The guide withholds memory stores for sessions on self-hosted environments. It also withholds the pre-built self-hosted sandbox workers in the ant CLI and the Python, TypeScript, and Go SDKs. For that case, the guide tells callers to implement their own worker that calls the Environments Work endpoints.
| Where it stands | What Anthropic states |
|---|---|
| Setup guide confirms these | Fast mode for Opus 5.5. Web search, web fetch, code execution, the memory tool, the advisor tool, and the MCP connector. Agent Skills and the Files API. |
| Setup guide confirms these | Prompt caching, extended thinking, server-side fallback, streaming, structured outputs, and token counting. The Message Batches API, the Models API, and a Compliance API toggle. CMEK with Google Cloud KMS. |
| Setup guide confirms these | Claude Managed Agents: agents, environments, sessions, credential vaults, memory stores, scheduled deployments, and self-hosted sandboxes. The blog marks Managed Agents, advisor strategy, and the MCP connector as beta. The guide does not repeat those labels. |
| Named on the blog, absent from the guide’s feature section | Computer use and citations. The guide says that section does not say whether an unlisted feature works. |
| Setup guide says these are not supported | API keys. MCP tunnels, separate from the supported MCP connector. Memory stores on self-hosted environments. The pre-built self-hosted sandbox workers in the ant CLI and the Python, TypeScript, and Go SDKs. The guide’s substitute is a worker you write that calls the Environments Work endpoints. |
Billing, retention, and the other Google Cloud path
The setup guide says these steps create a new Anthropic organization for the Google Cloud project, separate from any organization the company already has, including a Claude Enterprise organization. An existing Anthropic organization cannot be converted, including one used with the Claude API directly. API keys, workspaces, and Claude Console settings do not carry over. Files, skills, batches, and Claude Managed Agents resources stay in the existing organization.
The setup guide says Anthropic meters usage and reports that usage to Google Cloud in Claude Consumption Units. One unit is $0.01 USD. The charges appear on the Google Cloud bill, and Anthropic does not send a separate invoice. That unit is the Marketplace meter, separate from the fast-mode per-million token rates on the blog. If you have a Google Cloud spend commitment, the guide says usage counts toward it the way other Marketplace purchases do. If you already have a discount, contact your Anthropic or Google Cloud account representative before you start. The guide says a discount cannot be applied to past usage.
The setup guide says Anthropic is the data processor for this offering. Some supporting services, including Anthropic’s trust and safety systems, process data outside Google Cloud. The guide says the offering follows the same data retention policy as the Claude API. Zero Data Retention is available on request through an Anthropic account representative, and an arrangement on another Anthropic organization does not carry over. The announcement and the setup guide do not include that policy’s default window.
Anthropic’s blog describes two paths. It says the Anthropic-managed offering is for customers who want day-zero access to beta and generally available Claude Platform features, with Anthropic operating the service and acting as the data processor. It says the Google-managed offering remains the best choice when Google Cloud must be the data processor, when regional data residency requirements are strict, or when the customer wants native Google Cloud integrations such as Model Armor or Agent Engine. Anthropic manages rate limits on the Anthropic-managed offering. The announcement and the setup guide do not publish the numeric limits.
Source
Also from
Sources and editorial approach
This article relies on the ClaudeDevs post and the @claudeai post from 22 October 2026, Anthropic’s announcement, and the Claude Platform guide to Claude Models on Google Cloud (Anthropic managed). Those pages were read on 23 October 2026. The Claude Code changelog notes dated 22 October 2026, at the start of that page, do not mention this offering. The rest of the changelog was not read. Linked pages on the offering comparison, data residency, numeric rate limits, CMEK eligibility, context windows, and the Claude API’s default retention rules were not read. This article does not include an API test.