Google Cloud’s OpenAI Deal Lets Companies Use Existing Spending Commitments

OpenAI and Google Cloud announced October 21 that companies will be able to purchase OpenAI models through Google’s cloud marketplace and apply existing spending commitments to those purchases. Companies already buying computing from Google gain a purchasing route through an existing supplier. General availability is planned within weeks. OpenAI announcement

Google names three models in the initial private preview: GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. Selected enterprise customers are getting access first. Google Cloud announcement

Model Garden is Google’s catalog for finding, testing and deploying AI models. The catalog connects model selection with Google’s tools for evaluating and running them. Customers can compare options within a platform rather than beginning every search on a model developer’s website. Model Garden overview

Google’s general partner-model guide describes managed services with an API, a connection through which software sends requests. Customers do not need to provide the hardware. A service endpoint is the address software calls to request an answer. This arrangement lets application teams use a model without maintaining the machines serving it. The OpenAI preview’s own instructions will establish its supported connection method. Partner-model administration guide

For Google, selling OpenAI gives customers another reason to keep AI spending inside its cloud. For OpenAI, the arrangement offers access to companies that have already negotiated a Google contract. Those are commercial advantages each company can pursue even while its models compete for the same tasks.

A cloud commitment still costs money

Google says OpenAI usage will appear on one Google Cloud invoice, reduce existing cloud commitments and count toward its Flexible Savings Plans. Google Cloud announcement

Those savings plans involve a promise to spend a minimum amount each month for one or three years. Google’s documentation says third-party products receive no additional discount from the plan. It also says customers remain responsible for unused monthly commitments, which do not roll forward. A purchasing team should distinguish eligibility to consume a commitment from a discount on the model itself. Flexible Savings Plans documentation

That distinction changes which price a buyer should compare. Money already promised to a cloud supplier can help pay for eligible AI use, but it remains company spending. A team deciding between direct access and a marketplace needs both rates and its remaining commitment balance. Choosing a more expensive model merely to use a commitment can still raise the cost of completing a task.

OpenAI has used this purchasing approach before. Its Oracle partnership was announced in June; an August 7 update said customers could buy API access, Codex and ChatGPT Work through Oracle Marketplace using eligible Oracle Universal Credits. That update specified United States availability, with further regions planned. Google is adding another buying route rather than inventing cloud-funded AI purchasing. Oracle Marketplace announcement

What companies can test in the preview

OpenAI’s access form describes an invitation-only evaluation for approved development and internal uses. It lists the Responses API, which applications use to send work to a model and receive results. It also lists Codex access through its app, command-line tool and supported coding environments. These are specific ways to evaluate the integration, rather than an unrestricted production launch. OpenAI preview requirements

Google’s separate interest form says applying does not guarantee admission. A company preparing a trial should identify its proposed use, the staff who need access and the information they will send. That gives its technical and purchasing teams something concrete to review while access is being considered. Google Cloud preview interest form

Existing Google documentation separates permission to enable partner models from permission to send requests. An administrator must allow the relevant service and model under organization policies. Those general controls explain why a working Google account alone does not establish that an employee can activate every model in the catalog. Preview-specific instructions still need to come with the invitation. Partner-model administration guide

A useful comparison would run the same permitted tasks through the current model and a preview candidate, recording successful completions, response times and charges. Google’s evaluation service supports task-specific criteria, reference answers and comparisons using a company’s own examples. Its documentation also distinguishes interfaces: the console supports Gemini models, while the software kit can reach a broader set of models. A shared catalog does not make every evaluation interface identical. Model evaluation documentation

For coding work, measuring one response would miss time spent correcting errors or repeating an attempt. A buyer can record whether the change passes its existing tests, how many attempts it takes and how much review a developer performs. Those are proposed trial measures; the partnership announcement supplies no head-to-head results establishing which purchasing route produces cheaper completed work.

What Google promises about business data

Google says customer prompts, context and outputs stay within designated data boundaries and are not used to train foundation models. It describes authenticated identities and unified auditing for model interactions. Google Cloud announcement

In the same announcement, Google marks Zero Data Retention and Confidential Compute as forthcoming options. Google Cloud announcement

These features address different questions. OpenAI’s existing description of Zero Data Retention concerns what happens to prompts and responses after processing, subject to eligibility and stated exceptions. Google’s confidential-computing documentation describes hardware protections for information while it is being processed. A no-training policy, retention setting and protection during computation therefore require separate checks. OpenAI retention announcement Confidential Computing overview

Location needs similar care. Google’s endpoint documentation warns that choosing an endpoint does not by itself guarantee where data will be processed. Global routing can send requests to different supported regions, while supported multi-region configurations address broader jurisdictional boundaries. A company with a location requirement needs the commitments for the particular model and service, rather than borrowing assurances from another product’s region list. Google Cloud endpoint documentation

For a prospective buyer, an invitation is the time to obtain the model identifier, allowed uses, retention settings and applicable processing locations together. The same review should cover request limits, support and the actual billing rate. A successful internal experiment still needs an operating plan covering who receives failures, who can change settings and which team controls spending before it becomes an approved business application.

The distribution agreement also leaves an existing relationship in place: OpenAI says Microsoft remains its primary cloud partner. Google’s addition gives buyers another purchasing option within that wider arrangement. OpenAI announcement

OpenAI says general availability is planned in the coming weeks. The next useful documents will be the final access instructions, prices and supported configurations. Those will let purchasing teams compare what their contract pays for with what developers can actually run. OpenAI announcement