OpenAI Models Come to Google Cloud: What Is in Preview

Google Cloud's October 21, 2026 announcement puts it in one sentence: OpenAI's GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna "are rolling out now in private preview to select enterprise customers in Google Cloud's model garden, with more coming soon". OpenAI's matching post adds the two facts a buyer needs: "General availability is planned in the coming weeks", and "Microsoft remains OpenAI's primary cloud partner." For an enterprise with a Google Cloud commitment, the question is whether to request preview access now or wait for GA, and the answer depends on facts neither post states.

The context to keep in view is that this is the third of the three largest clouds to sell GPT-6, not the first. Microsoft Foundry announced GPT-6 Astra as generally available on September 3, 2026 and Amazon Bedrock on September 8, and both list GPT-6.1 Sol and GPT-6 Luna as well. So the news is what Google Cloud customers gain and when, and what Google offers that the other two do not, which is Gemini on the same platform. Sources were read on October 21, 2026 (CEST); every status label is the cloud's own wording.

Google's post is written by Michael Gerstenhaber, its VP of product management for Gemini, and it is the more specific of the two. It names the three models and describes each in a line: Astra for demanding work across coding, computer use, professional tasks and scientific research; 6.1 Sol as near-Astra intelligence at lower cost; Luna as the fast, cost-efficient model for everyday work at scale. It states the access route, an interest form, and the purchase route, Google Cloud Marketplace. It promises that OpenAI models will come to Gemini for business and Google Antigravity "coming soon", with no plan tiers or dates. OpenAI's post names no models at all; it describes "frontier models and Codex agentic workflows", confirms the Marketplace and commitment-drawdown mechanics, and gives the GA timing.

The example of why the wording matters: "private preview to select enterprise customers" means a sales conversation, not a console toggle. The failure case is a team that assumes it can enable the model today the way it enables a Gemini or Claude model. When we read Google's partner-model documentation on October 21, the page had been updated that day and did not list any GPT-6 model, and the documentation path for OpenAI partner models returned a 404. The announcement is ahead of the catalogue, which is normal for a private preview and is the reason the GA date matters.

The date record is short and it comes from the other two clouds' own pages. Microsoft's Azure blog of September 3, 2026 says GPT-6 Astra "is now generally available for all customers in Microsoft Foundry", and its September 22 post adds Sol and Luna to the generally available line-up, with GPT-6.1 Sol following on September 29. AWS's What's New post of September 8, 2026 opens with "Today, AWS announces the general availability of GPT-6 Astra from OpenAI on Amazon Bedrock", and Bedrock's model cards list GPT-6.1 Sol (launched September 29) and GPT-6 Luna (September 22) as active with Marketplace product identifiers. Both clouds also still list the earlier GPT-6 Sol, which Google does not name.

That sequence is why this post is framed as what Google Cloud customers gain rather than as a first. The gain is real: an enterprise with a Google Cloud commitment could not previously buy GPT-6 through it, and the same enterprise cannot buy Gemini through AWS, which lists only Gemma among Google's models, or through Azure, which lists no Google model. The failure case is the opposite inference, that Google's route is the mature one. On October 21 it was the only one of the three in private preview and the only one without a published rate card. The full three-cloud matrix, with every status label and date, is in which cloud sells which model.

Billing is the most concrete part of the announcement, and it is the reason the launch matters to procurement rather than to engineers. Google's sentence: "Because OpenAI models are transacted directly through Google Cloud Marketplace, all usage appears on a single Google Cloud invoice, draws down existing cloud commitments, and counts toward Flexible Savings Plans (FSP) across model providers." OpenAI's version says enterprises with Google Cloud spend commitments will be able to draw down those commitments for these purchases. Taken together: no separate OpenAI contract, no second invoice, and spend that counts against a commitment the company has already signed.

The detail that is easy to misread is the FSP line. Google's own Flexible Savings Plan documentation, read the same day, says the plan covers participating third-party models through the Marketplace, that spend on them draws down against the enterprise commitment under the standard Marketplace rules and contract caps, and that "third-party products don't get any discounts from the FSP". So the benefit of buying GPT-6 through Google is drawdown, not a cheaper token. What the token costs on this route is not published anywhere we could find; AWS and Azure both publish GPT-6 rate cards, and the comparison belongs in our frontier model price index once Google states a number. Our earlier piece on using cloud commitments for AI spend explains the drawdown mechanics on the other two clouds.

The one capability both posts put at the centre is Codex. Google says that "starting today, customers can build with these models through our APIs and connect their local Codex CLI and Desktop workflows directly to our Gemini platform for inference". OpenAI describes the partnership as bringing its frontier models "and Codex agentic workflows" into the model garden. In practice this means a developer running Codex on a laptop can point its inference at a Google Cloud project, so the tokens are billed and governed there rather than through an OpenAI account.

The example that shows why an enterprise cares: a bank with a Google Cloud perimeter can let engineers use Codex without the model calls leaving that perimeter or appearing on a corporate card. The failure case is treating this as Google-only. AWS's September 8 post said customers could configure ChatGPT Work and Codex to use the model on Bedrock, and Bedrock's FAQ lists managed agents powered by OpenAI in preview. Codex-on-cloud is a feature of OpenAI's distribution strategy rather than a Google exclusive, and the choice between them comes down to where the commitment and the data already live, the same decision we worked through for Claude Code across Bedrock, Vertex and Foundry.

Google's post groups its security statements under three headings, and they split into present and future tense. Present: customer prompts, context and outputs "remain entirely within your designated data perimeter and are never used to train foundation models"; queries are routed through hardened, isolated gateways so that third-party model providers have no unauthorised access to the customer's cloud environment; and every interaction runs under an authenticated identity with VPC Service Controls protection and unified audit logging. Future: zero data retention options are "coming soon", and OpenAI models run on dedicated Google Cloud infrastructure "with Confidential Compute coming soon".

For a regulated buyer the two future-tense items are the decision. If your policy requires ZDR for any third-party model, the Google route does not meet it on October 21, while AWS's documentation says eligible customers may request full ZDR for OpenAI models through their account team. The failure case is reading "dedicated infrastructure" as confidential computing; Google separates the two in the same sentence. The example to run is a one-line check with your Google account team: when ZDR arrives, will it apply per project or per organisation, and will it cover the preview period retroactively? Neither post says.

Google's headline claim is quoted above, and the three-cloud table lets a reader check it. On the catalogue pages we read on October 21: AWS Bedrock lists the GPT-6 family and the Claude 5.5 family as generally available and no Gemini model, only Gemma. Microsoft Foundry lists the GPT-6 family and the Claude 5.5 family and no Gemini model. Google Cloud lists Gemini 4 Argon and the Claude 5.5 family as generally available, and the GPT-6 family in private preview per the blog, with no public catalogue entry yet.

So the claim holds as a statement about access, with one qualification Google does not make: it is the only one of the three whose platform includes Gemini, and it is also the only one where GPT-6 is not yet generally available. "Access to all three" is true in the preview sense. A buyer who needs all three in production today has it on no single cloud; a buyer who needs Gemini plus Claude in production and GPT-6 in weeks has it on Google Cloud alone. That is the honest version of the headline, and it is still a meaningful one.

Six things a buyer would ask for are absent, and each is a reason the preview label matters. A price per token on the Google route. Regions. Model identifiers and quotas, which would normally appear on the catalogue page that does not yet exist. A GA date, as opposed to "weeks". Whether the earlier GPT-6 Sol, which AWS and Azure still sell, will be offered. And the terms under which the preview runs, which for a private preview on Google Cloud are normally pre-GA terms with no SLA. We looked for each on both vendors' documentation and found none; we will update this page as they appear rather than estimate any of them.

One further gap is on OpenAI's side. Its post names no models, and its "Microsoft remains OpenAI's primary cloud partner" line is a statement about priority. Azure's own September post went further, saying OpenAI "continues to launch first on Azure". The practical reading for a Google Cloud buyer is that new OpenAI models should be expected on Google after Azure and probably after AWS, as this launch itself demonstrates.

The routing below assumes the facts as of October 21 and will be wrong the day Google publishes a rate card and regions. The principle behind it is simple: a private preview is for learning, and a commitment drawdown is for production, and the two should not be confused in the same quarter's budget.

Where the three clouds differ on model availability, billing and data handling is a table rather than a paragraph, and we maintain it as a reference. For organisations choosing a cloud route for frontier models under an existing commitment, our AI transformation team runs the evaluation and the procurement comparison together.