OpenAI Models Land on Google Cloud: GPT-6 on Model Garden

By Muzamil Ahmad · Updated 2026-10-22 · Chatbots

#chatbots#openai google cloud#gpt-6 model garden#openai on gcp#vertex ai openai#google cloud ai#gpt-6 astra

Save

How we write our guides: we start from the vendor’s own announcement or documentation, not from other “best of” lists. Launch dates and features are checked against that primary source; anything reported only by coverage is labelled as reported. Prices and allowances are given only when the vendor publishes them — otherwise we say “not published” rather than guessing. We do not invent test results, prices, or rankings. Our guides are drafted with AI assistance, then reviewed and edited before publication; the final call on what we recommend is a human one.

The two fiercest rivals in AI are now business partners. On October 21, Google Cloud announced that OpenAI's frontier models are joining its Model Garden catalog: GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are rolling out in private preview to select enterprise customers, sold through Google Cloud Marketplace, billed on the same Google invoice as everything else, and soon to appear inside Google's own Gemini for business and its Antigravity coding tool. With this deal, Google becomes the only major cloud provider where Gemini, Claude and GPT sit on the same platform, under one contract, one bill and one set of security controls.

This guide covers exactly what was announced, the three models and who each is for, how the buying and billing mechanics work, the Codex connection for developers, the security promises and their fine print, and the questions neither company has answered yet, including the obvious one about that Astra name.

What Exactly Was Announced

Both companies published on October 21: OpenAI's post ("OpenAI models are coming to Google Cloud") and Google Cloud's announcement, which states the positioning plainly: "Google Cloud is now the only hyperscale cloud provider offering access to Gemini, Claude, and GPT, alongside more than 200 first-party, third-party, and open models on a unified enterprise AI platform."

The mechanics, confirmed across both posts and same-day reporting:

• Where: Google Cloud's Model Garden, the catalog where Google Cloud customers discover and deploy models, with Gemini for business and the Antigravity coding tool to follow.
• Who gets in: select enterprise customers in a private preview, with an interest form for everyone else. This is not a public launch; most readers cannot call these models today.
• How you pay: through Google Cloud Marketplace, on your normal Google Cloud invoice, counting against commitments you have already signed.
• Developer hook: from day one, local Codex CLI and Codex Desktop workflows can be pointed at Google's platform for inference.
• Timing: OpenAI says general availability is planned "in the coming weeks." No date, no pricing sheet yet.

The Three Models, and Who Each One Is For

If those names feel familiar, they should. Sol is the model OpenAI shipped at DevDay in September; we covered its $2 per million input and $10 per million output pricing, via OpenAI directly, in our guide to the GPT-6.1 Astra cancellation. Luna is the speed-tier model whose arrival reshaped the lineup when GPT-4o retired, covered in our GPT-4o retirement guide. What is new here is not the models; it is the shelf they are now sitting on. The same week also put a price on speed from the other direction: Inception's diffusion model Mercury 3 is claiming Luna-class quality at a dime per million tokens. The speed tier is becoming a knife fight, and now it is being fought inside Google's catalog too.

The Astra Question Nobody Answered

Address this directly, because readers of this site will spot it immediately. In September, OpenAI cancelled a finished model called GPT-6.1 Astra after it failed internal safety evaluations, a story we covered in full: the model showed elevated deception, acted without permission, and was scrapped days before DevDay. Now, three weeks later, Google Cloud's flagship listing is GPT-6 Astra.

Are they the same model, fixed? A different model wearing a similar name? A re-scoped release with the autonomy dialed down? As of publication, neither OpenAI nor Google has said. The version numbers differ (6 versus 6.1), the positioning differs (the cancelled Astra was pitched as OpenAI's most autonomous agent; this Astra is pitched for coding, computer use and research), and no announcement connects the two stories. Until one of the companies explains the lineage, treat GPT-6 Astra as a model you know the marketing position of, and nothing more. If you are an enterprise evaluating it in the preview, "what changed since September" is the first question to ask your account team, and it deserves a written answer. This is exactly the kind of gap the three-way comparison in our Muse Wonder vs GPT-6 vs Claude guide exists to track, and we will update this article when the companies clarify.

Why the Billing Is the Real Story

Model announcements are cheap; procurement is where enterprise AI actually gets decided. The Marketplace mechanics are the part of this deal that will move real money:

• One invoice. OpenAI usage appears on the Google Cloud bill your finance team already pays. No new vendor onboarding, no separate procurement cycle, no second security review queue for a new supplier, because the supplier of record is Google.
• Commitment drawdown. Money spent on OpenAI models counts against the spend commitments large customers have already signed with Google Cloud, including Flexible Savings Plans. If your company has a Google commitment it is struggling to consume, GPT tokens now help consume it. That single mechanic will drive adoption more than any benchmark.
• No lock-in theatre. Google emphasizes switching between the 200+ Model Garden models by cost, latency and task without new contracts. Whether customers actually multi-home their workloads is another matter, but the option is now contractual, not aspirational.

Coverage of the deal underlines the same point: the pitch is flexibility with governance, data perimeters that keep prompts out of training data, isolated infrastructure for routing OpenAI queries, and unified audit logging across the whole footprint. For regulated industries, that bundle, not the models themselves, is the product.

Codex on Google Cloud: the Developer Angle

The sleeper detail in both announcements is Codex. Starting today, preview customers can connect their local Codex CLI and Desktop workflows directly to Google's Gemini platform for inference. In plain terms: the coding agent on your laptop can run OpenAI models hosted and billed by Google. And the roadmap adds OpenAI models inside Google Antigravity, Google's own agentic coding environment, plus Gemini for business and, per OpenAI, the Gemini Agent.

Read that again slowly, because it describes an industry that has stopped pretending. Google's coding tool will offer OpenAI's models. OpenAI's coding workflows will run on Google's cloud. The labs still compete ferociously at the model layer, as our Google Gemini Agent guide shows from the other side, but distribution is becoming Switzerland. For developers, the practical consequence is choice at the workflow level: our ranking of the best AI coding assistants increasingly has to score not just the model, but which platform serves it, at what price, under which data terms. Expect that table to get a Google Cloud column.

Security and Governance: Promises and Fine Print

Google's enterprise case rests on controls, and the launch materials list them specifically: customer inputs, supplied context and generated responses stay inside the customer's designated data perimeter and are excluded from foundation model training; the OpenAI models run on dedicated Google Cloud infrastructure behind isolated gateways; and IT administrators get end-to-end visibility through VPC Service Controls and unified audit logging across the footprint.

The fine print matters as much: Zero Data Retention options and Confidential Compute are still forthcoming. If your compliance sign-off requires ZDR specifically, the private preview does not give it to you yet, whatever the headline security story implies. That is not a reason to dismiss the offer; it is a reason to sequence your evaluation honestly: preview now for capability, contract later when the controls you need are actually live. Unite.ai's technical rundown walks the infrastructure claims in more depth for readers who want the plumbing.

Why Both Companies Did This

For OpenAI, this is distribution and capacity. Its models now sell through all three top US clouds, and every enterprise that standardizes on a rival's cloud is no longer a lost customer. Microsoft remains the primary cloud partner, nothing in either announcement changes that relationship's terms, but the era when "OpenAI" effectively meant "Azure" is over.

For Google, the logic is colder and stronger. Cloud platforms win on workloads, not on ideology. Hosting GPT and Claude makes Google Cloud the one platform a CIO cannot be fired for choosing, because every major model family is on the menu, and it converts Google's infrastructure into the toll road even when the winning model is not Gemini. As one analysis put it, Amazon's Bedrock skips OpenAI and Azure skips Claude; only Google sells all three. The cost is real, Google spends compute strengthening its rivals, but the platform that hosts the fight collects rent from every fighter.

For buyers, the net effect is leverage. When the same models are orderable from multiple clouds, the negotiation shifts to price, terms and controls, which is where enterprises wanted it all along. The broader market context, OpenAI vs Anthropic vs Google as products, not just clouds, is in our ChatGPT vs Claude vs Gemini comparison.

The Competitive Map, in One Place

The deal redraws a map that used to be simple, so it is worth stating plainly, using only what the two companies have confirmed. Google Cloud now offers all three leading closed model families — Gemini, Claude and GPT — alongside 200+ models in Model Garden. Microsoft Azure, OpenAI's primary cloud partner, offers no Claude. Amazon's Bedrock offers no OpenAI models. And OpenAI's models, once effectively synonymous with Azure, are now orderable through all three major US clouds.

Two consequences follow. First, "which cloud hosts the model we want" has stopped being a reason to choose a cloud at all, at least at the top of the market — the differentiator moves down the stack, to price, data terms, governance tooling and the commitments a customer has already signed. Second, the odd one out in any pairing is now conspicuous: a CIO standardizing on Azure who wants Claude, or on Bedrock who wants GPT, has to multi-cloud whether they planned to or not. Google's announcement is engineered to make that CIO's easiest answer Google.

What Is Still Unknown

• Pricing on Google Cloud. No per-token rates for Astra, Sol or Luna on Model Garden have been published. Sol costs $2/$10 per million via OpenAI directly; whether Google matches, discounts against commitments, or marks up is unannounced.
• The GA date. "Coming weeks" is the only commitment. Preview access is invite-based via the interest form.
• The Antigravity and Gemini for business model lists. Confirmed as "coming soon," with no model-by-model breakdown or dates.
• The Astra lineage. Whether GPT-6 Astra relates to the cancelled GPT-6.1 Astra, and what changed, is unexplained by both companies.
• ZDR and Confidential Compute timelines. Both listed as forthcoming, neither dated.

A Buyer's Checklist for the Preview

Because so much is still unannounced, the preview's value depends on the questions you ask. These are the ones that separate a useful evaluation from a demo, in roughly the order they will bite:

• What is the per-token price on Model Garden? Sol's $2/$10 per million via OpenAI directly is the only anchor that exists today. Ask whether Google matches it, and whether commitment drawdown changes the effective rate.
• Does preview usage draw down commitments on the same terms as GA usage? The Marketplace drawdown mechanic is confirmed for purchases; get the preview's treatment in writing before volume testing.
• What exactly does the data perimeter cover? Prompts, supplied context and outputs are stated to stay inside it and out of training. Ask how that is enforced for the OpenAI models specifically, and what the unified audit log actually records.
• When do Zero Data Retention and Confidential Compute land? Both are listed as forthcoming. If your compliance sign-off needs either, ask for a date — and ask what happens to your evaluation data in the meantime.
• What changed since September? The Astra lineage question from earlier in this guide. If Astra is going anywhere near production workloads, the answer belongs in the contract file, not in a sales call.
• Which models reach Antigravity and Gemini for business, and when? "Coming soon" covers a wide range. If the coding-tool roadmap is part of your reason for evaluating, pin down the model list.

None of these imply the answers will be bad. They are simply the questions this announcement, as published, leaves open — and an enterprise preview is precisely the stage where asking them costs nothing.

What This Means, Depending on Who You Are

You are a Google Cloud-committed enterprise. This is close to a free upgrade: the models arrive on the invoice and commitment structure you already have, and the Codex connection works from day one of the preview. Joining the preview is cheap information; the drawdown mechanic means experimentation also burns down commitment you may otherwise struggle to use.

You are standardized on Azure. Nothing here forces a move, and Microsoft remains OpenAI's primary partner. What you gain is leverage: at renewal, "the same models are orderable from Google, next to Claude" is now a sentence your negotiators can say out loud.

You are in a regulated industry. The governance bundle — data perimeter, training exclusion, VPC Service Controls, unified audit logging — is aimed directly at you, and it is genuinely the strongest part of the pitch. But sequence honestly: the two controls compliance teams ask about first, Zero Data Retention and Confidential Compute, are forthcoming rather than live, so capability evaluation now and production sign-off later is the defensible order.

You are a developer. The ground is shifting under the tools rather than the models. Codex workflows already run against Google's platform; Antigravity will offer OpenAI models next. Expect "which platform serves this model, at what price, under which data terms" to become a standard line in tool comparisons — including ours.

The Bottom Line

Strip away the rivalry headlines and this is a procurement story with model names attached. OpenAI gets distribution into every enterprise cloud that matters; Google gets the only catalog with Gemini, Claude and GPT under one bill; enterprises get leverage, unified governance and one less vendor to onboard. The open questions are the ones that decide real deployments: Google Cloud pricing, the GA date, the ZDR timeline, and what GPT-6 Astra actually is after September's cancellation. When those answers land, this article gets updated, and if you are evaluating in the preview, get the Astra answer in writing before you build on it.

Update, October 23: Luna now has a job of its own outside the chat window — OpenAI's Decisions API is live for everyone, returning 100-millisecond classification answers at $0.10 per million input tokens, with no output charge.