Google Cloud announced on October 21, 2026, that OpenAI models are coming to its Model Garden, giving selected enterprise customers another way to access frontier AI models through Google’s cloud platform. The announcement includes GPT-Astra, GPT-6.1 Sol, and GPT-6 Luna, as well as support for connecting local Codex CLI and Codex Desktop workflows to Google’s platform for model inference.
The practical change is about distribution and enterprise integration—not a claim that every Google Cloud customer can immediately use every OpenAI model. Google says the models are rolling out in private preview to selected enterprise customers. Organizations interested in using them should request access and confirm eligibility, supported configurations, regional availability, quotas, and commercial terms before planning a production deployment.
The primary source is the official Google Cloud announcement. Reuters also reported on the partnership and its enterprise distribution implications in its coverage of OpenAI models coming to Google Cloud.
Which OpenAI models are included in Google Cloud Model Garden?
Google’s October 21 announcement identifies three models and describes them as options for different types of enterprise work.
GPT-Astra: demanding professional and technical tasks
Google positions GPT-Astra for demanding workloads that can include coding, computer use, professional tasks, and scientific research. Teams considering it should evaluate the model against their own task requirements, such as multi-step reasoning, software engineering, tool use, reliability, and response latency.
The announcement does not mean GPT-Astra is the right default for every workload. High-capability models can be more than a task requires, so teams should compare quality and total cost with alternatives before routing all traffic to a single model.
GPT-6.1 Sol: agentic coding and professional workflows
Google describes GPT-6.1 Sol as offering intelligence close to Astra at lower cost across agentic coding, computer use, and professional work. That positioning makes it worth evaluating for workflows that require an AI model to take several steps, use tools, or work with code.
Actual cost and performance depend on the model configuration, usage pattern, and applicable commercial terms. The announcement alone does not establish a universal price advantage for every application.
GPT-6 Luna: efficient work at scale
GPT-6 Luna is presented as a fast, cost-efficient option for everyday work at scale. It may be a candidate for high-volume tasks where latency and cost matter, provided that testing shows it meets the application’s quality and reliability requirements.
The right selection process is workload-specific: test a representative sample, measure output quality and failure rates, record latency, and compare costs before making a model the default.
What changes for Codex CLI and Codex Desktop?
One of the announcement’s most relevant details for developers is that they can connect local Codex CLI and Codex Desktop workflows to Google’s platform for inference, subject to the announced access and setup requirements.
Codex is an AI coding agent, while the model provides capabilities used during a task. Bringing a model into a cloud platform does not automatically mean every Codex installation is configured for that model or that every user has access. Teams should follow Google’s current onboarding instructions and confirm which accounts, client versions, and model options are supported.
For a development team, the potential benefit is the ability to evaluate OpenAI models within an existing Google Cloud environment rather than treating model selection, cloud procurement, and governance as entirely separate decisions.
Before connecting a local workflow, administrators should verify authentication, access permissions, billing ownership, logging, and the data that coding tools may send with requests. Repository contents can include proprietary code, internal documentation, secrets, and customer information; use established secret-scanning and access-control practices rather than assuming that model access alone secures the development workflow.
If you are comparing coding agents, see our guide to Claude Code vs. Codex vs. Gemini CLI. For broader Google Cloud agent workflows, see Google Cloud’s Gemini Agent and enterprise workflows.
How private preview affects availability
Google says GPT-Astra, GPT-6.1 Sol, and GPT-6 Luna are rolling out in private preview to selected enterprise customers in Model Garden. Private preview is a restricted access stage; it is not general availability.
Organizations should therefore take the following steps before treating the integration as an available production dependency:
- Request access. Use the interest or onboarding process linked from the official Google Cloud announcement.
- Confirm eligibility. Ask whether the organization, billing account, region, and intended workload qualify for the preview.
- Check supported models and interfaces. Verify the specific model identifiers, API or client support, quotas, and any restrictions that apply to the preview.
- Review commercial terms. Confirm model pricing, billing mechanics, usage limits, and whether the intended usage qualifies for existing cloud commitments.
- Run a controlled pilot. Start with a bounded workload and non-sensitive test data. Measure quality, latency, reliability, and cost before expanding access.
- Plan for change. Preview capabilities and terms may change before general availability. Avoid designing a critical production dependency around an unconfirmed feature or service-level commitment.
Google also says OpenAI models are expected to become available in Gemini for business and Google Antigravity in the future. Those are described as coming-soon integrations in the announcement, not as generally available features today.
Enterprise data governance and security: what Google says
Google’s announcement emphasizes using its existing enterprise controls with the new model options. It says customer prompts, context, and outputs remain within the customer’s designated data perimeter and are not used to train foundation models. The announcement also describes authenticated identities, unified audit logging, and VPC Service Controls protection across the enterprise environment.
Some details are explicitly described as forthcoming. In particular, Google says Zero Data Retention (ZDR) options and Confidential Compute are coming soon. Organizations should not interpret those future items as already enabled for every preview customer.
Security and compliance teams should validate the exact terms and technical behavior applicable to their own account. Important questions include:
- Which regions and services are in scope for the data perimeter?
- What request and response data is retained, and for how long?
- Which identities can invoke each model, and how are permissions reviewed?
- What events appear in audit logs, and who can access those logs?
- Which controls are available during private preview, and which are still planned?
- How are secrets, customer records, and regulated data excluded from prompts where required?
For teams operating agents in production, our guide to AI agent observability: traces, metrics, logs, and privacy explains why monitoring, workflow traces, cost controls, and privacy-aware telemetry should be designed into the system.
Billing through Google Cloud Marketplace
Google says OpenAI model usage will be transacted through Google Cloud Marketplace. According to the announcement, usage will appear on a single Google Cloud invoice, can draw down existing cloud commitments, and counts toward Flexible Savings Plans (FSP) across model providers.
That could simplify procurement and cost management for organizations that already buy substantial Google Cloud services. However, the financial effect depends on each customer’s contract, eligible commitments, usage volume, and the specific terms applied to the preview. It should not be treated as a guaranteed discount or as evidence that the new route is cheaper than buying access elsewhere.
A practical cost comparison should include model usage, request volume, any related infrastructure charges, the cost of moving data, operational overhead, and the value of consolidating procurement. Compare equivalent workloads and quality requirements rather than comparing headline prices in isolation.
How to evaluate OpenAI models on Google Cloud
A small, structured pilot can help teams decide whether this integration is useful for their environment.
1. Choose a representative task
Select a bounded use case such as code review, internal knowledge retrieval, document analysis, or a multi-step agent workflow. Avoid starting with a critical production process or unrestricted access to sensitive systems.
2. Establish a baseline
Record the current model’s output quality, task completion rate, latency, failure modes, and cost. Use the same evaluation set when comparing the OpenAI models available through Model Garden.
3. Compare models on the same workload
Test GPT-Astra, GPT-6.1 Sol, and GPT-6 Luna only if access is granted and the models are supported for the intended use. Track quality and cost together. A lower-cost model is useful only if it still meets the application’s accuracy and reliability requirements.
4. Test the full workflow, not just the answer
For coding and agentic tasks, evaluate tool selection, file changes, command execution, permission handling, and recovery from errors. Require review for consequential changes and retain logs that help explain failures.
5. Validate data and access controls
Check authentication, permissions, logging, retention, and data handling with the security team. Do not assume that all planned controls are available during private preview; verify the current configuration and contractual terms.
6. Define a rollback path
Keep a tested alternative model or workflow available where practical. Preview services can change, and a production design should not rely on unconfirmed availability or undocumented behavior.
Does this make Google Cloud a multi-model AI platform for enterprises?
The announcement strengthens Google’s effort to offer enterprises a choice of models and infrastructure through a unified platform. Organizations may value being able to evaluate Google models and third-party models within a common environment, particularly when governance, billing, and deployment processes are already built around Google Cloud.
But model availability alone does not guarantee portability. Differences in APIs, tool calling, context limits, model behavior, rate limits, and safety features can require application changes when switching models. Teams should use abstraction layers carefully and maintain model-specific tests rather than assuming that every model is interchangeable.
Google’s announcement frames this as a way to choose the best model for each workload based on capability, latency, and cost. The useful enterprise question is whether that flexibility reduces friction without adding unacceptable complexity to security, evaluation, and operations.
What is available now—and what is coming later?
As of Google’s October 21, 2026 announcement:
- Rolling out now: GPT-Astra, GPT-6.1 Sol, and GPT-6 Luna in private preview for selected enterprise customers in Model Garden.
- Announced developer workflow: Connecting local Codex CLI and Codex Desktop workflows to the Google platform for inference, subject to access and setup requirements.
- Coming soon: OpenAI models in Gemini for business and Google Antigravity.
- Coming soon: Zero Data Retention options and Confidential Compute, as described by Google.
- Not established by this announcement: General availability for all Google Cloud customers, universal access across regions, or a guaranteed cost reduction for every workload.
Because the integration is in private preview, customers should rely on the latest Google Cloud instructions and their own account’s access status before making implementation commitments.
Bottom line
Google Cloud’s October 21 announcement brings GPT-Astra, GPT-6.1 Sol, GPT-6 Luna, and Codex-related workflows into a new enterprise distribution channel. The potential value is a combination of model choice, existing cloud governance, consolidated billing, and the ability to evaluate OpenAI models in an environment enterprises may already use.
The limitation is availability: the three models are rolling out to selected enterprise customers in private preview, while some integrations and security options remain forthcoming. The best next step is to request access, verify the current terms, and run a measured pilot before making the integration a production dependency.
Official sources
- Google Cloud: OpenAI models are coming to Model Garden
- Reuters: OpenAI to sell its top models on Google Cloud
- Google Cloud Model Garden