OpenAI on Google Cloud: Models, Codex Access and Private Preview Details

OpenAI and Google Cloud announced on October 21 that OpenAI models and Codex workflows are coming to Google's Model Garden. Selected enterprise customers can enter a private preview now, while general availability is planned for the coming weeks. For existing Google Cloud customers, the immediate attraction is a simpler route to evaluating another model family.

Which OpenAI models are available?

Google's announcement names OpenAI's Astra model, GPT-6.1 Sol and GPT-6 Luna for the private preview. It says selected customers can use APIs and connect local Codex CLI and desktop workflows to Google's platform for inference.

This is a restricted rollout. A public announcement does not mean every Google Cloud account can immediately enable these models, and the interest form is not a guarantee of admission.

The announcements do not provide a complete public pricing table, regional availability matrix or universal quota schedule. Those details should be confirmed before putting a production workload behind the preview.

What changes for purchasing and access?

OpenAI says purchases will go through Google Cloud Marketplace, with eligible purchases able to draw down customers' GCP commitments. It also says Microsoft remains its primary cloud partner.

The useful distinction is between model capability and the route through which a company buys and operates that capability. An organization may already have a good model evaluation process but still face a separate procurement exercise before it can use the preferred model. This partnership could reduce that friction for teams already committed to Google Cloud.

That does not establish that the Google route is cheaper. A commitment can change where spending is allocated without changing the cost of each completed task. An evaluation should include the full workflow: retries, tool calls, cached input, output length and operational overhead.

Google Cloud governance has staged availability

Google describes dedicated Google Cloud infrastructure, identity and audit controls, and a designated data perimeter. It says customer prompts, context and outputs are not used to train foundation models. However, Zero Data Retention options and Confidential Compute are described as coming soon.

Treat current controls and planned controls as separate requirements. A security review that needs one of the planned capabilities cannot assume the preview already satisfies it.

For example, a team evaluating an internal code-review assistant could begin with a non-sensitive test repository. It could compare the same changes across models, inspect logging behavior and check whether the identity attached to each request matches its access policy. That gives the team concrete evidence before it considers a broader rollout.

OpenAI in Antigravity is still coming

Google says access through Gemini for business and Google Antigravity is coming soon. OpenAI also describes future availability in Antigravity and the Gemini Agent. These announcements should not be read as confirmation that OpenAI models are already generally selectable in those products.

There is a practical reason to keep these surfaces distinct. API access, a coding client's model picker and an enterprise agent product can have different rollout schedules and administrative requirements. A successful API request would not establish that every related interface has shipped.

What should developers compare?

A useful pilot would keep the task, available tools and acceptance criteria fixed. For coding work, that could mean asking each model to resolve the same small set of repository issues, then comparing passing tests, review findings and total cost rather than just the first answer.

The hosting route also deserves its own check. Confirm which model identifier is actually used, where requests are processed, how failures are reported and what happens when quota is exhausted. Model choice is only useful if the application can observe and manage those differences.

I think the most interesting implication is that enterprise model selection can become less tied to a single purchasing channel. But that only helps when teams retain portable evaluations and clear operational requirements. A larger model menu should produce better evidence for a decision, not a longer list of models nobody has tested.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams turn complex launches and product decisions into clear stories that developers and founders can use.

Book a 30-minute call