OpenAI models join Google Cloud’s menu, but access is still invite-only

The change is where a team can discover and buy the models

OpenAI announced a Google Cloud partnership on October 21 to bring its frontier models and Codex agentic workflows into Google Cloud’s Model Garden. The pitch is familiar to companies that already run workloads in Google Cloud: find and test another model in the same cloud environment, then connect it to existing software and data workflows. OpenAI says purchases will eventually run through Google Cloud Marketplace.

The arrangement also changes how customers can pay. OpenAI says Google Cloud customers with qualifying spend commitments will be able to draw down those commitments for model purchases after the offer becomes generally available. This is a purchasing route, not a promise of a lower price. The post does not include a public price sheet, a model-by-model catalog, or the detailed rules for which customer agreements qualify.

The current offer is a limited preview, not an open catalog

OpenAI’s request-access page marks the offer “Available in limited preview.” It says selected customers can evaluate OpenAI frontier models through the Responses API on Google Cloud for supported development and internal testing use cases. Codex is available through its API in the Codex app, command-line interface, and supported development environments. The preview is invite-only; submitting the form is a request to be considered, not confirmation that access has been granted.

The request form also asks for a work email, company, role, country or region, Google Cloud relationship, and primary use case. Applicants must confirm that the private preview will be used for approved development and internal purposes. That makes the access boundary fairly clear: a business can ask to participate, but it should not design a production migration around the model before it has been accepted, provisioned, and given concrete service terms.

Model Garden, Codex, and Gemini integrations have different status

OpenAI’s post says general availability is planned for the coming weeks, and the request page still calls the Model Garden and Codex access a private preview. The announcement also says OpenAI models will soon appear in Google Antigravity and Gemini Agent. These are separate routes: Model Garden access, Codex through its supported development clients, and future use from Google’s own agent environments. A line about a planned integration should not be mistaken for a link that is already enabled in every account.

Axios’ October 21 report describes another route to customers of the three major U.S. cloud providers: businesses could use a competitor’s model without moving their whole workflow to a different cloud. It places the agreement after OpenAI’s earlier expansion to Amazon Web Services and notes that Microsoft remains OpenAI’s primary cloud partner, as OpenAI’s own post confirms. The Google agreement expands distribution and procurement choices rather than announcing an exclusive switch. A company can use different providers for different workloads; access through Google Cloud does not make OpenAI’s models part of Google’s Gemini model family.

The cloud environment does not settle every data question

OpenAI says the integration will pair its models with Google Cloud’s access controls, centralized governance, and infrastructure. Those are important platform features, but the announcement does not identify the data region for each model, explain request retention, list logging defaults, or say which company’s terms govern every part of an API call. A Google Cloud project can provide useful identity and billing controls while the model provider still has separate service settings and contractual commitments.

The practical questions therefore remain account-specific. Before sending confidential material through a preview endpoint, a team should confirm the available region, the model and endpoint identifier, what logs are kept, whether prompts or outputs are used for improvement, which support staff can access them, and how usage is billed. The public announcement does not answer those questions; the preview agreement and service documentation should.

A Marketplace line item can make procurement easier to track, but it does not answer where requests are processed, what endpoint stores logs, or which agreement governs support access. The preview form's limit to approved development and internal testing is a use restriction, not a statement that production is already supported. A team should get those answers from its preview terms before it sends a real customer record or source-code secret.

Codex is a workflow route, not just another model card

The Codex part of the announcement is aimed at software development. OpenAI says teams will be able to use Codex through its API in the Codex app, CLI, and supported IDEs, while the Model Garden route exposes frontier models through Google Cloud. That may let a company evaluate a model and a coding workflow in an environment where its repositories and cloud services already operate. It does not mean a Codex client automatically gains access to every Google Cloud resource.

A coding agent needs its own credentials and tool permissions. A team should test it first on a bounded repository, with read access separated from permission to modify branches, open pull requests, or touch deployment credentials. The customer remains responsible for deciding what the agent can reach; a cloud partnership does not make a code-review or approval policy unnecessary.

A team should also check which development surface it is actually being invited to use. The announcement names the Codex app, CLI, and supported IDEs for the preview; it does not promise that every feature in the existing Codex service, every repository connection, or every organization policy transfers unchanged to the Google Cloud route. Developers should follow the named preview path and verify how identity, repository access, and tool permissions are configured there.

Use the preview to answer one deployment question at a time

A useful evaluation could compare an existing endpoint with the Google Cloud preview on the same approved workload. Record the model version, region, input, tools, latency, output quality, and billable usage. For a coding task, keep the repository and test environment fixed and measure whether the change passes tests and how much human correction it needs. Do not infer price or performance from a marketplace route alone.

Then test the part the partnership is meant to change: can developers reach the model without exporting workflow data to an unfamiliar platform, can administrators apply their existing identity policies, and can finance teams reconcile spending through their normal cloud account? The current preview gives selected customers a way to investigate these questions. Until OpenAI and Google publish general availability details, the reliable description is limited access now, wider distribution planned.