On October 21, 2026, OpenAI and Google Cloud announced a partnership to offer OpenAI's latest AI models in Google Cloud's "Model Garden." The three models are GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna, and a private preview for select enterprise customers has already begun. OpenAI models are also planned to come to Google's AI development environment, "Antigravity."
For companies that use Google Cloud, the partnership does more than add AI model options. They can run OpenAI models on Google Cloud infrastructure and apply the usage fees toward spending under their existing cloud contracts. This lets enterprises choose the model developer, the cloud platform that actually runs the model, and the party they pay separately. However, the enterprise preview and the Antigravity rollout are not starting at the same time, and for data protection features, it is necessary to distinguish between what is available now and what will come later.
Three models on Google Cloud, and plans for Antigravity
Model Garden is Google Cloud's model catalog, where users can search for and select AI models for their needs. Google Cloud describes the three OpenAI models being added as follows.
| Model | Main uses and features as described by Google Cloud |
|---|---|
| GPT-6 Astra | Handles coding, computer use, and professional work, as well as advanced tasks such as scientific research |
| GPT-6.1 Sol | Offers capabilities close to Astra at lower cost, supporting agent-based development and professional work |
| GPT-6 Luna | Suited to processing large volumes of routine work quickly and at low cost |
These are, however, Google Cloud's own characterizations of each model, not the results of a performance comparison under identical conditions. The announcement does not give per-model processing speeds, benchmark results, or specific pricing for using the models through Google Cloud. It is therefore difficult to judge, from the descriptions alone, which model is best suited to a given business or which would be the cheapest to run.
The timing also differs. Google Cloud has started a private preview for some enterprise customers, and OpenAI says general availability is planned within the next few weeks. The addition of OpenAI models to Antigravity, by contrast, has been announced by both companies only as a future plan.
The timing of general availability on Google Cloud and the timing of availability in Antigravity should therefore be considered separately. The announcement also does not reveal which specific models will be added to Antigravity, what will be available under individual plans, or what limits such as usage caps will apply. Features offered first to enterprises will not necessarily be opened to individuals on the same terms.
OpenAI model fees folded into existing Google Cloud contracts
Another important aspect of the partnership is that the tools developers work with can be separated from the infrastructure that runs model inference. According to Google Cloud, customers in the preview can use OpenAI models through an API, and can also connect Codex CLI and Desktop, which run in a local environment, to the company's Gemini platform and run inference on Google Cloud. Developers can keep using familiar tools while moving the model execution platform to Google Cloud. The models run not on a local computer but on dedicated computing infrastructure provided by Google Cloud.
Payment is handled through Google Cloud Marketplace, and OpenAI model usage fees are consolidated into the Google Cloud invoice. Usage can also be applied toward cloud spending commitments made in existing enterprise contracts. Google Cloud says it will also support "Flexible Savings Plans (FSP)," which cover multiple model providers.
FSP itself is not a new program created for this partnership. In a billing program announced on August 26, 2026, Google Cloud described a mechanism in which customers commit in advance to a monthly spending amount. The program offers discounts of 10% for a one-year commitment and 20% for a three-year commitment, and Google Cloud says FSP spending can also count toward existing Google Cloud enterprise agreements.
The partnership does not merely add OpenAI models to Google Cloud's model catalog; it also builds them into existing billing and budget management. Companies that already have a Google Cloud contract may be able to manage fees for new AI models within that contract, reducing the burden of adding payees and budget management systems for each model. This is a particular advantage for companies that are comparing multiple AI models as they adopt them, since procurement and cost management can be unified.
However, it cannot be assumed that FSP discount rates apply directly to the effective price of each OpenAI model. The announcement does not make clear whether unit prices are the same when using models directly from OpenAI and through Google Cloud, or how much discount applies under individual contracts. Being able to consolidate billing and budget management is separate from usage fees necessarily becoming cheaper.
Not using data for AI training is not the same as not retaining it
When companies adopt AI models, how input data is handled is an important consideration alongside price and performance. Google Cloud says that customer prompts, contextual information, and model outputs will be handled within the specified data governance boundary and will not be used to train foundation models. It also says existing Google Cloud security features can be used, including access control through authenticated identities, protection with VPC Service Controls, and centralized audit logs.
However, a promise not to use input data for model training is not the same as a mechanism that does not retain data. In Google Cloud's October 21 announcement, not using customer data to train foundation models is described as part of what is currently offered, while "Zero Data Retention (ZDR)," which does not retain data, and "Confidential Compute" are positioned as features to be offered in the future.
| Data management / security feature | Availability as of October 21, 2026 |
|---|---|
| Prompts, contextual information, and outputs not used to train foundation models | Described as currently offered |
| Execution on Google Cloud's dedicated infrastructure, authentication and access control, audit logs | Described as currently offered |
| Zero Data Retention (ZDR): an option that does not retain data | Planned for the future |
| Confidential Compute | Planned for the future |
This difference is especially important for companies that handle confidential information. A statement that data will not be used for training defines the purposes for which the service provider may use the data; it does not mean the retention period is zero. Companies whose internal rules or laws set strict data retention conditions therefore need to confirm that the necessary retention conditions are met at the time of adoption, in addition to being able to use existing authentication and audit features.
Google Cloud also says that requests to the models are processed through an isolated gateway, but this does not mean Confidential Compute is already available. The announcement also does not give specific launch dates or supported regions for ZDR and Confidential Compute. The range of security features companies can actually use needs to be confirmed in terms of both the service's implementation status and contract terms.
More freedom in model choice, but operating conditions must be checked
The partnership lets companies combine AI models, the cloud infrastructure that runs inference, and the tools developers use more flexibly than before. Running OpenAI model inference on Google Cloud while using Codex is one concrete example. If the rollout to Antigravity proceeds, it will also add a route for selecting OpenAI models from Google's development environment.
If multiple AI models can be evaluated under the same billing system and identity management, companies should find it easier to avoid setting up separate procurement and audit procedures each time they compare models. Whether models can actually be switched without changing the application, and how much work migration would take, is a separate question. The announcement includes no measured data showing how much development effort or operating cost would be saved by switching models.
OpenAI also stated in the announcement that Microsoft remains its primary cloud partner. The expansion to Google Cloud does not end the relationship with Microsoft; it can be seen as an effort to add routes for companies to use OpenAI models from their existing cloud environments.
Going forward, the key factors for companies deciding whether to adopt will be the specific per-model prices through Google Cloud and when the necessary data protection features will become available. If companies can evaluate the models using existing cloud budgets and also meet requirements such as data retention conditions and usage regions, OpenAI models will be easier to build into their current operating environments. For Antigravity, meanwhile, information needed to begin an actual evaluation, such as supported models and terms of use, is still awaiting future announcements.