Google Cloud announced on Wednesday, October 21, that it is adding OpenAI's GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna models to Model Garden. The rollout starts with a private preview for selected enterprise customers, who receive access to the models through Google Cloud infrastructure and application programming interfaces.
The expansion concerns the development of business applications and AI agents. Google provides an interest form for API access, requesting information about the company and its contact person. Participation involves selection for the private programme.
Three models for different types of work
Model Garden is Google's catalogue for discovering, testing, customising and deploying AI models. It brings together Google's own models, partner models and open-weight models. The platform integrates with evaluation and model-serving tools, while each model card describes the available options for using that model.
OpenAI positions GPT-6 Astra as its model for the most demanding work. Its public model card lists complex reasoning, coding, computer use, research and document creation. On the OpenAI API, the model accepts text and images as input and produces text. The documentation specifies a context window of 1,050,000 tokens and a maximum output of 128,000 tokens.
OpenAI positions GPT-6.1 Sol for complex work at a lower cost than Astra. The company describes its performance as close to Astra's in coding, computer use and professional tasks. The API model card lists five reasoning effort settings, ranging from low to max, with medium as the default.
GPT-6 Luna is the option OpenAI positions for focused tasks and high volumes of requests. Its documentation includes a mode without additional reasoning, alongside configurable effort settings. Like the other two models, it accepts text and images and returns text. These specifications are described in the API model cards published by OpenAI.
Google's catalogue supports filtering by task, provider, collection and features. The same interface includes models with different deployment arrangements: some have managed APIs, while others are deployed within the customer's cloud project. The new programme for OpenAI models is being introduced as API access for approved participants.
Usage is billed through Google Cloud Marketplace
According to Google, transactions for these OpenAI models take place through Google Cloud Marketplace. Usage appears on the Google Cloud invoice, can draw down existing cloud commitments and counts towards Flexible Savings Plans. The company is linking model selection to the commercial arrangements already used for its enterprise infrastructure and services.
Flexible Savings Plans are based on monthly spending commitments. In its August presentation of the programme, Google described one-year and three-year terms and the ability to use commitments within existing enterprise agreements. The October 21 announcement brings usage of the new OpenAI models into this commercial framework alongside other participating providers.
Model Garden documentation distinguishes between accessing a model as a service and deploying a model within a customer's own cloud project. For customer-deployed models, infrastructure management and computing resources form part of the deployment. In the new programme, Google describes OpenAI models running on dedicated Google Cloud infrastructure, with requests reaching them through the platform's interfaces.
Identity, access policies and company data
In Model Garden, organisation policies can control access to specific models at the organisation, folder or project level. Administrators can authorise models their company has assessed and deny access to others. This mechanism forms part of catalogue administration, alongside the technical model cards and deployment options available for each model.
For applications involving agents, Gemini Enterprise Agent Platform includes individual identities, a registry of agents and tools, and a central point for enforcing policies. Agent Identity manages an agent's identity, while Agent Gateway controls authentication and rules for tool calls. The platform documentation also covers behaviour evaluation, monitoring and execution tracing.
In its description of the new integration, Google says customer prompts, context and outputs remain within the designated data perimeter and are not used to train foundation models. The company lists authenticated identities, VPC Service Controls protection and unified audit logs for interactions with agents and models.
Google lists Zero Data Retention options and Confidential Compute among the features to be added later. Their status is separate from the private API access being introduced now. For the initial infrastructure, the company describes hardened, isolated gateways that route requests to the OpenAI models.
Developer access and announced expansions
Selected participants can use the models through Google Cloud APIs and connect their local Codex CLI and Codex Desktop workflows to the Gemini platform for inference, according to the announcement. Inference is the stage at which a model processes a request and produces a response, separate from model training. It is the computing component covered by the connection from development tools.
Google also included a perspective from Box through Yashodha Bhavnani, its vice president of AI products. She described using OpenAI models through Google Cloud as a way to combine access to models with security and compliance capabilities for enterprise content. The statement describes the direction of Box's product and its integration plans.
The next announced expansions are the availability of OpenAI models in Gemini for business and Google Antigravity. In the phase that began on October 21, the access communicated by Google is through APIs and connected development workflows, within the private programme for selected enterprise customers.
