OpenAI's Frontier Models Land on Google Cloud in New Deal

OpenAI's frontier models are now running on Google Cloud. On October 21, 2026, OpenAI and Google Cloud said that GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna have entered private preview inside Google Cloud's Model Garden, with Codex agentic workflows included, billing routed through Google Cloud Marketplace, and general availability promised "in the coming weeks." It is a notable crossing of a competitive line: OpenAI's own models, sold through the cloud infrastructure of one of its fiercest rivals in both AI models (Gemini) and cloud computing.

What OpenAI and Google Cloud actually announced

The announcement landed on both companies' own sites on the same day. OpenAI published a short post titled "OpenAI models are coming to Google Cloud," confirming that its frontier models and Codex agentic workflows will be brought into Google Cloud's Model Garden, the catalog Google uses to let enterprise customers discover, test and deploy first-party, partner and open models inside Vertex AI.

Google Cloud's side of the announcement came from a company blog post, "Google Cloud brings OpenAI models to customers," written by Michael Gerstenhaber, Google Cloud's VP of Product Management for Gemini. The post frames the deal in terms of choice: Google Cloud says it is now the only hyperscale cloud provider offering customers access to Gemini, Claude and GPT models side by side, among more than 200 models already listed in Model Garden.

That framing matters. Google has spent the past two years building Model Garden into a multi-vendor storefront rather than a Gemini-only showcase, and Anthropic's Claude models have already been available there through Vertex AI. Adding OpenAI's models completes a set: all three of the industry's most-discussed frontier model families, bought and billed through one cloud console.

Which OpenAI models showed up in Model Garden

Google Cloud's blog post names three models and describes how it expects customers to use each one:

Google's pitch is that customers should match the model to the workload rather than defaulting to one model for everything, and that switching between Astra, Sol and Luna — or between those and Gemini or Claude models already in Model Garden — shouldn't require a separate vendor contract or a rearchitected pipeline.

Notably absent from the list is GPT-6.1 Terra, the model OpenAI released days after the Google Cloud announcement. Terra is not part of the Model Garden preview as of this writing, which suggests OpenAI is staggering which models it distributes through Google's infrastructure rather than mirroring its entire current lineup on day one.

Codex workflows come along for the ride

The deal isn't limited to raw model access. Both companies' posts describe Codex, OpenAI's agentic coding product, as part of the package. Google Cloud's blog says customers in the preview can "connect their local Codex CLI and Desktop workflows directly to our Gemini platform for inference" — meaning a developer's existing Codex setup can route its model calls through Google Cloud rather than OpenAI's own infrastructure, while the tooling itself stays the same.

Neither company's post includes setup instructions, specific latency numbers or benchmark comparisons for running Codex against Google's infrastructure versus OpenAI's own, so those details remain unconfirmed pending wider rollout.

Pricing and billing: one invoice, existing commitments

Both companies emphasize that this is a commercial integration, not just a technical one. Per Google Cloud's post, OpenAI models purchased through Model Garden are transacted directly through Google Cloud Marketplace, which means:

• Usage shows up on a single Google Cloud invoice alongside other cloud spend.
• Purchases can draw down a customer's existing Google Cloud spending commitments.
• Usage counts toward Flexible Savings Plans (FSP) that span multiple model providers, not just Google's own models.
• Customers don't need to negotiate or sign a separate vendor contract with OpenAI to buy its models this way.

OpenAI's own announcement echoes this, telling enterprise customers that organizations can apply their existing Google Cloud spending commitments to OpenAI model usage bought this way. Neither company has published specific per-token pricing for Astra, Sol or Luna on Google Cloud as of this article, and no discount or markup relative to OpenAI's direct API pricing has been disclosed by either source.

Private preview now, general availability still pending

As of publication, OpenAI's models on Google Cloud remain in private preview, limited to select enterprise customers. Google Cloud's blog post says customers can already build against the APIs and points interested organizations to a sign-up form to request access, but it stops short of naming a general-availability date. OpenAI's own post goes slightly further, telling readers that broader, general availability is planned "in the coming weeks" without committing to a specific date.

Both companies describe further expansion as in motion but undated: Google Cloud's post lists "more OpenAI models" as coming soon, along with availability inside Gemini for business and Google Antigravity, Google's agentic development environment. Neither post has been updated with a GA date since the original October 21 publication, based on the versions live at the time of writing.

What's missing: data retention and confidential compute

Enterprise buyers evaluating this deal will likely focus as much on governance as on model quality. Google Cloud's post is explicit that, for now:

• Prompts, context and outputs stay inside the customer's designated Google Cloud data perimeter and are not used to train foundation models.
• OpenAI models run on dedicated Google Cloud infrastructure, with queries routed through hardened, isolated gateways so that data in transit stays inside Google Cloud's own environment.
• Every model and agent interaction runs under an authenticated identity, with VPC Service Controls and unified audit logging layered on top.
• Zero Data Retention (ZDR) options are listed as "coming soon" — not yet available in the preview.
• Confidential Compute is likewise listed as "coming soon," on what Google describes as a separate timetable from the private preview itself.

In other words, the governance controls that heavily regulated customers (finance, healthcare, government contractors) typically require before adopting a third-party model aren't fully in place yet. Google Cloud's post includes one customer quote addressing this directly: Yashodha Bhavnani, VP of AI Products at Box, said that being able to use OpenAI's models through Google Cloud will let Box "give customers access to OpenAI's frontier models alongside Google Cloud's security and compliance capabilities," calling it "a step toward operating more efficiently as we bring AI to customers."

Why this crosses a real rivalry line

The reason this deal reads as more than a routine cloud-distribution announcement is who's involved. Google and OpenAI compete directly on frontier models — Gemini versus GPT — and increasingly on developer tooling, where Google's own coding agent plugged into Gemini CLI has already been pulled into a separate Antigravity CLI tool, a product line that competes head-on with Codex. Selling OpenAI's models through Google's cloud, rather than exclusively through Microsoft Azure or OpenAI's own API, means Google now profits from infrastructure spend on a rival's models even as it sells its own competing models in the very same catalog.

OpenAI's motivation looks straightforwardly practical: more places to run inference, and access to enterprise customers who are already committed to Google Cloud and would rather not sign a separate OpenAI contract to use GPT-6 models. Both of OpenAI's own posts on this deal are explicit that Microsoft remains OpenAI's primary cloud partner — this is additive distribution, not a replacement of that relationship. The move sits alongside a broader pattern of OpenAI diversifying its compute and distribution footprint beyond a single partner, at a moment when demand for inference capacity across the industry continues to outstrip the capacity any single cloud provider can offer.

It also lands in a market where competitive and financial stakes for AI labs are unusually visible: Anthropic, OpenAI's other major rival, has recently drawn scrutiny over security findings disclosed in its own IPO filing, a reminder that distribution deals like this one are being struck against a backdrop of heavy investor and regulatory attention on how frontier models are deployed and secured.

How this fits OpenAI's release calendar

The Google Cloud preview lands in the middle of a dense stretch of OpenAI model releases. GPT-6 Astra, Sol and Luna were already part of OpenAI's lineup before this announcement; GPT-6.1 Terra, which OpenAI shipped only days after the Google Cloud news broke, is a newer, separate release that is not currently included in the Model Garden preview. That gap is worth watching: if OpenAI brings Terra into Model Garden in a later wave, it would suggest the company plans to keep Google Cloud roughly in step with its main model lineup rather than treating it as a secondary, lagging distribution channel.

For now, enterprise customers evaluating Astra, Sol or Luna on Google Cloud should treat the preview as functionally equivalent to using those same models directly through OpenAI's API, with the main practical differences being procurement (Marketplace billing, existing GCP commitments) and infrastructure (Google's data centers and network, rather than OpenAI's or Microsoft Azure's), not model capability itself — neither company's announcement claims any difference in model behavior or output quality based on which cloud it runs on.

What it means for enterprise AI buyers

For a company already running workloads on Google Cloud, the practical upside is procurement simplicity rather than anything new in model capability. Being able to pull GPT-6 Astra, Sol or Luna into an existing Vertex AI pipeline, bill it against a Google Cloud commitment already baked into a budget, and manage access through the same identity and audit tooling used for every other cloud service removes a category of friction that otherwise requires setting up a separate OpenAI enterprise account, a separate contract, and a separate line item for finance to track. That is the same procurement logic that made Anthropic's Claude models attractive inside Vertex AI before this announcement, and Google Cloud's post leans on that precedent directly when it describes itself as the only hyperscaler offering Gemini, Claude and GPT side by side.

The tradeoff, at least for the length of the private preview, is that the compliance story is incomplete. A regulated buyer that specifically needs Zero Data Retention or Confidential Compute guarantees before it will send production traffic to a third-party model cannot yet get those guarantees for OpenAI models running on Google Cloud, even though the same buyer might already have those protections for its Gemini or Claude usage inside the same console. Until Google Cloud ships both features for this integration, enterprises with the strictest data-handling requirements are likely to stay on OpenAI's direct API or on Microsoft Azure's existing OpenAI Service, where that groundwork has had years longer to mature.

What's next

Three things to watch in the coming weeks, per both companies' own statements: a general-availability date that moves this out of private preview and opens it beyond hand-picked enterprise accounts; the arrival of Zero Data Retention and Confidential Compute, which Google Cloud has explicitly flagged as still pending; and the promised extension of OpenAI models into Gemini for business and Google Antigravity, which would put GPT-6 models inside Google's own agentic products rather than only inside the neutral Model Garden catalog. None of those have a confirmed date as of this writing, and this article will be updated if either company publishes one.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage