Google Cloud adds OpenAI GPT-6 models to Model Garden in private preview

Google Cloud and OpenAI announced on October 21, 2026, that OpenAI’s newest models are rolling out in Google Cloud’s Model Garden, the catalogue of first-party, partner and open models on Google’s enterprise AI platform. The most important fact for builders: this is a private preview for select enterprise customers, with general availability planned “in the coming weeks,” according to OpenAI.

Three models are named for the first wave: GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. Usage is bought through Google Cloud Marketplace and lands on a Google Cloud invoice. With this deal, OpenAI’s models are now offered on the platforms of the three leading cloud infrastructure providers.

What Google Cloud and OpenAI announced on October 21

Google Cloud says it is rolling out OpenAI’s frontier models in Model Garden, and OpenAI says its latest models and Codex agentic workflows are coming to Google’s cloud. Codex is OpenAI’s coding agent.

Google frames the addition as a platform milestone. It says it is “now the only hyperscale cloud provider offering access to all three leading closed source LLMs,” meaning Gemini, Anthropic’s Claude and OpenAI’s GPT, alongside more than 200 first-party, third-party and open models on one enterprise platform.

Google Cloud CEO Thomas Kurian described the goal in customer terms:

“We are focused on giving customers a broad selection of models, including those from OpenAI, so they can choose the right model for any given job, helping to deliver optimal quality and lower costs.”

OpenAI’s Greg Brockman pointed to greater choice for customers and growth of the agent ecosystem, according to one report. OpenAI’s own announcement closes with a line aimed at its existing partner: “Microsoft remains OpenAI’s primary cloud partner.”

The three OpenAI models in the first preview

Google describes each model by the work it is meant for. These are Google’s product descriptions, not independent benchmarks.

Model	What Google says it is for	Reported relative price
GPT-6 Astra	Flagship: demanding coding, computer use, professional work, scientific research	Premium frontier tier
GPT-6.1 Sol	Intelligence close to Astra at lower cost, for agentic coding, computer use and professional work	Roughly one-fifth of Astra’s cost, as reported by SiliconANGLE
GPT-6 Luna	Fast, cost-efficient everyday work at scale	Roughly one-hundredth of Astra’s cost, as reported by SiliconANGLE

Two caveats apply to that table. First, Google has not published its own per-model prices in the announcement, so the relative figures are a press report and billed prices should be treated as unconfirmed until an account team quotes them. Second, outlets render the names slightly differently: some write “GPT-Astra” and “GPT 6.0 Luna”, while others use GPT-6 Astra and GPT-6 Luna. Teams should wait for the exact model IDs in Google’s documentation before writing code.

On safety, OpenAI’s own notes describe Astra as producing “fewer misaligned outcomes” than the other frontier models it evaluated, with configuration options that stop it from bypassing user-defined task restrictions, according to one summary of the announcement. That is OpenAI’s evaluation, not a third-party result. Unicorn API has previously covered GPT-6 Astra in a report on rival labs and the model.

Google also says more OpenAI models will be added to the library later.

Who can use it now, and what is still coming

Access is the part most likely to be misread. The announcement does not mean every Google Cloud project can call these models today. Here is the status as the sources describe it:

• Google Cloud APIs: preview customers can build with the models through Google’s APIs now.
• Codex CLI and desktop app: preview customers can route inference from local Codex command-line and desktop workflows into Google’s Gemini platform. This is a preview route, not something live in every Codex install.
• General availability: planned in the coming weeks per OpenAI; no firm date has been given.
• Google Antigravity and Gemini Agent: listed as coming soon. Antigravity is Google’s agent-powered coding tool.

The two companies name the future Gemini integration differently: OpenAI says Gemini Agent, while Google says Gemini for business. Gemini Agent was launched on October 8 at “Gemini at Work 2026” as a general-purpose work agent spanning Workspace, Microsoft 365 and Slack, designed to route each task to the most suitable model; at launch it could route between Gemini and Claude, with OpenAI and open-weight models listed as “coming next,” according to IT之家.

Not every outlet reports the status the same way. One article says the integration “officially launched” on October 21 with “full API compatibility”. Other coverage, citing Google’s announcement, describes a limited private preview for select enterprise customers, and that is the status this article reports.

It also helps to separate this news from what already exists. Google’s documentation lists OpenAI’s older gpt-oss open models separately from partner models, so an OpenAI name in Model Garden does not prove access to the new models. For general Model Garden work, the gcloud command-line tool already offers commands to list publisher models, list supported machine configurations and deploy a model to a Vertex AI endpoint. No source says those commands are the route for the OpenAI preview.

Billing runs through Google Cloud Marketplace

The buying mechanics are central to the deal. Model Garden hosts the models; Google Cloud Marketplace handles the purchase. According to Google:

• OpenAI model usage appears on a single Google Cloud invoice.
• Eligible customers can apply existing Google Cloud spending commitments to those purchases.
• Usage counts toward Flexible Savings Plans across model providers.
• No separate agreement with OpenAI is needed.

Google also says customers can select and swap models for any workload based on performance, latency and cost, without separate vendor contracts.

Paying through existing commitments changes how an organisation pays, not what inference costs. Actual savings depend on commercial terms, model pricing, usage volume and existing commitments.

For Google, the commercial logic is that it captures cloud spending even when a customer picks another company’s model. Google’s cloud business has grown quickly this year: annualised revenue growth rose to about 82% in the second quarter from 63% in the first, compared with 37% for Amazon and 42% for Microsoft in the second quarter, according to Deep News.

Data handling claims, and where reports differ

Google says prompts, supplied context and generated responses stay inside the customer’s designated data perimeter and are excluded from foundation model training. The models run on dedicated Google Cloud infrastructure, with requests passing through isolated gateways. Administrators get unified audit logs and VPC Service Controls protection, a Google Cloud feature for fencing off resources inside a security boundary.

Reports differ on two higher-assurance options. One outlet says Zero Data Retention (ZDR), meaning requests are not stored after processing, and Confidential Compute were announced. Another says both are still listed as forthcoming rather than available at launch. Teams with strict data rules should confirm which controls apply to their preview before sending sensitive data.

All of these are the provider’s claims, not an independent audit.

What it means for builders

For teams already running on Google Cloud, the practical change is that they can test GPT models without a new contract or billing arrangement, once they have access. Developers can use existing Google Cloud patterns, including IAM permissions, VPC networking and audit logging, according to one outlet. This mirrors a broader trend toward vendor-neutral tooling, such as Mistral’s Responses API preview in OpenAI’s format.

The bigger picture is distribution. OpenAI’s exclusive agreement with Microsoft expired in April 2026, and Amazon announced the next day that it would host OpenAI models via AWS Bedrock; Google Cloud is the third provider. A team can now reach OpenAI models through more than one cloud and compare them side by side with Gemini and Claude on one platform. Unicorn API recently covered Google’s Gemini 4 Carbon, one of the Google models that will share that catalogue.

Before wider release, one guide suggests these steps:

1. Ask the cloud account owner whether your organisation is eligible for preview access.
2. Pick one representative workload and define acceptable quality, latency and budget.
3. List what to check once documentation appears: model IDs, regions, quotas, IAM roles and prices.
4. Confirm with procurement whether your billing account and spend deal cover the Marketplace purchase.
5. Record a baseline from your current model so any switch is measured and reversible.

Open questions remain. General availability, regions and pricing are not yet specified. Neither company has published a setup guide for all users or a public price list.

What to watch next

• General availability: OpenAI’s “coming weeks” timeline, and whether launch documentation lists model IDs, regions and quotas.
• Pricing: whether Google publishes per-model prices that match the relative figures SiliconANGLE reported.
• Data controls: when Zero Data Retention and Confidential Compute become available, given the conflicting reports.
• Antigravity and Gemini Agent: whether the integrations ship, and whether “Gemini Agent” and “Gemini for business” refer to the same product.
• More models: Google says further OpenAI models will join the library.
• Earnings: Google, Amazon and Microsoft are due to report quarterly results next week, according to Deep News, which may show how cloud AI demand is moving.