OpenAI Brings GPT-6 Models and Codex to Google Cloud’s Model Garden

Last updated: 23 October 2026. Figures and quotes below are sourced to primary publications linked inline.

OpenAI’s three paid flagship models are now sitting inside the cloud catalogue of one of its biggest frontier rivals. On 21 October 2026, OpenAI said that GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, along with Codex agentic workflows, had entered private preview inside Google Cloud’s Model Garden — the shelf Google also uses to sell its own Gemini models and Anthropic’s Claude. It is the first time OpenAI’s closed, frontier-grade models have been resold through Google’s cloud, and it puts the three best-known commercial model families in one enterprise shopping basket for the first time.

The 60-second version

• OpenAI’s GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna and Codex workflows are now in private preview on Google Cloud’s Model Garden, confirmed by OpenAI’s own announcement and matched by same-day reporting from The Next Web and Techzine.
• This is not OpenAI’s first rival cloud. Microsoft Azure’s exclusive hosting arrangement ended in April 2026, after which Amazon Bedrock and Oracle Cloud both added OpenAI model access. Google Cloud is the third addition and the first to also carry Google’s own competing Gemini line.
• No per-token pricing has been published. Enterprises buy through the Google Cloud Marketplace on a single invoice, and usage can draw down existing Google Cloud spending commitments — but Google’s own Flexible Savings Plan terms say third-party models don’t receive FSP discounts.
• Access is private preview only, limited to selected enterprise customers. OpenAI says general availability is coming “in the coming weeks,” with no firm date given.
• Google now says it is the only hyperscaler selling all three leading closed-source model families — its own claim, not an independently audited one.

Key dates and numbers

What actually happened

Three OpenAI model tiers are involved, according to matching accounts from The Next Web and Techzine: GPT-6 Astra for heavy workloads such as coding, computer use and scientific research; GPT-6.1 Sol, positioned close to Astra’s capability at a lower cost; and GPT-6 Luna, a fast, low-cost model aimed at high-volume everyday work. Codex, OpenAI’s coding-agent product, can already route inference through Google’s infrastructure, with support for Google Antigravity and a Gemini-side agent surface described as coming later — though the two outlets use different names for that surface, which suggests the integration itself is still being finalised.

OpenAI president Greg Brockman framed the move, in the company’s own announcement, as “making our most capable models and Codex available where customers are already building.” Google Cloud CEO Thomas Kurian said the goal was giving customers “a broad selection of models, including those from OpenAI,” and Google Cloud VP of Product Management Michael Gerstenhaber told Techzine the point was to let “each workload be matched to the right model, rather than relying on a single model for every task.” Google says prompts, context and outputs stay inside the customer’s data perimeter and are not used for training; protections such as Zero Data Retention and Confidential Compute are described as following shortly, which means they are not yet confirmed as live in the current preview.

None of this is OpenAI’s first appearance outside Microsoft’s cloud. Azure’s exclusivity arrangement ended in April 2026, and OpenAI model access followed on Amazon Bedrock that same month. OpenAI also announced an Oracle Cloud partnership in June 2026, though that listing wasn’t confirmed as actually purchasable until an August update — and, as of that update, only for customers in the United States, with other regions still pending. What is new is that Google — which is simultaneously trying to beat OpenAI at the frontier with its own Gemini 4 Argon model — is now also the company selling OpenAI’s models to its own enterprise customers, alongside Anthropic’s Claude.

Why a competitor is selling its rival’s models

The straightforward reading is procurement convenience: a Google Cloud customer who wants to compare GPT-6 Astra against Gemini or Claude on a real workload no longer needs a separate vendor contract, security review and invoice to do it. That is a genuine win for buyers, and it is consistent with how cloud marketplaces have worked for years — Google says Model Garden already lists more than 200 models in total across its own, open-source and third-party catalogue.

The more interesting reading is about where the actual leverage in frontier AI now sits. Distribution — who sells access to a model, and on whose invoice — has clearly gone multi-cloud: three of the four major US hyperscalers now sell OpenAI access, and Google sells OpenAI, Anthropic and its own Gemini side by side. Compute — the data-centre capacity that trains and serves these models — has not become multi-cloud in the same way. It remains locked into long-term, often exclusive arrangements: Microsoft is still described by The Next Web as OpenAI’s primary cloud partner for training and inference, and Techzine notes that Google’s own compute relationship with OpenAI, agreed in 2025, is separate from this Model Garden listing. OpenAI’s own balance sheet makes the stakes of that compute dependency obvious — its $30bn funding round earlier this month valued the company at $1.4 trillion, a figure that reportedly tracks its compute bill almost exactly. Anthropic’s infrastructure manoeuvring tells a similar story from the other side: its $35bn Lambda deal turned Nvidia into something closer to a landlord than a chip supplier, underlining how much of the real competitive battle in this industry is being fought over who controls the data centres, not who appears on a model-selection dropdown.

So when Google says it is “the only hyperscale cloud provider offering access to all three leading closed source LLMs,” that is a true statement about its shelf, not a claim about who controls the underlying silicon and power. Cloud providers typically take a commission on model sales made through their marketplaces, which is a reasonable commercial incentive for Google to host a rival’s models even while competing with them directly for the same enterprise customers.

What most coverage is getting wrong — or leaving out

A lot of the initial coverage leaned on the novelty of “the big three in one place” without flagging the financial fine print: Google’s own Flexible Savings Plan documentation states that third-party products do not receive FSP discounts. That matters because enterprises negotiate those committed-spend discounts specifically to lower their effective cloud costs, and a customer who assumes GPT-6 Astra usage will be discounted the same way their Gemini usage is could be in for an unpleasant finance-team conversation. Coverage has also been inconsistent on when OpenAI models first reached a non-Microsoft cloud — The Next Web dates the Amazon Bedrock listing to April 2026, while Techzine describes it as “the summer of 2025.” That discrepancy between two otherwise careful outlets is a useful reminder to check a primary announcement where one exists, rather than trusting secondary dates alone.

It is also worth noting that model-routing behaviour — letting an application switch between model providers rather than committing to one — is already a mature pattern in the industry, not something this announcement invents. Stripe’s acquisition of OpenRouter, a platform built entirely around routing requests across model providers at a reported 10 trillion tokens a day, shows there is already serious commercial infrastructure for exactly this kind of multi-model buying. Google’s Model Garden listing is a hyperscaler-native version of a problem independent routing platforms were already solving.

Practical takeaways for builders and publishers

For engineering teams already on Google Cloud, this is worth testing now, within its limits. You can get GPT-6 Astra, Sol or Luna onto the same invoice as your existing Google Cloud spend, which removes a genuine procurement bottleneck for running side-by-side model evaluations. But do not assume your negotiated Flexible Savings Plan rate applies to that usage — check your contract terms, because Google’s own documentation says it does not apply to third-party models by default.

For publishing and content operations evaluating model quality or cost for drafting, summarising or editing workflows, this is a reasonable moment to benchmark GPT-6 Luna’s lower-cost tier against Gemini or Claude equivalents already available on the same platform — without signing a new vendor agreement to do it. Just don’t build anything that depends on Zero Data Retention or Confidential Compute guarantees for OpenAI models on Google Cloud yet; both outlets reporting on the launch describe those protections as “following shortly,” which is not the same as live today.

For developers building on Codex specifically: CLI and desktop connections to Google’s infrastructure are available now, but deeper integration with Google’s own agent surfaces is explicitly described as coming later, and the two outlets covering the launch don’t even agree on what to call that surface yet. Treat anything beyond basic inference routing as unstable until OpenAI or Google publish their own documentation for it.

What we still don’t know

Several material details were not published alongside the announcement. There is no confirmed general-availability date — OpenAI’s own language is “coming weeks,” which is not a commitment. There is no published per-token pricing for any of the three models on Google Cloud, so cost comparisons against direct OpenAI API access or against Gemini and Claude pricing on the same platform are not yet possible. It isn’t clear what commission, if any, Google takes on sales made through the Marketplace listing, or whether that commission differs from what Google charges other third-party model providers already in Model Garden. The relationship between this distribution deal and Google’s separate 2025 compute arrangement with OpenAI — mentioned by Techzine but not detailed by either outlet — is also unclear: it is not obvious whether the two deals are commercially linked or entirely independent. And, as noted above, there is a genuine discrepancy in secondary reporting over when OpenAI’s models first reached Amazon’s cloud.

FAQ

Is this OpenAI’s first time on a cloud other than Microsoft’s?

No. Microsoft’s exclusive hosting arrangement ended in April 2026, and OpenAI model access subsequently appeared on Amazon Bedrock and Oracle Cloud before this Google Cloud listing. Google Cloud is the third added, and the first to host OpenAI’s models alongside both Anthropic’s Claude and Google’s own Gemini.

Which OpenAI models are available on Google Cloud?

GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, plus Codex agentic coding workflows, all currently in private preview for select enterprise customers.

Can I access these models on Google Cloud today?

Only if you are among the enterprise customers given private-preview access. OpenAI has not given a firm date for general availability.

Will using OpenAI’s models on Google Cloud cost less than using them directly?

There’s no published pricing yet to compare, and Google’s own Flexible Savings Plan terms state that third-party models, which would include OpenAI’s, do not receive the discounts applied to committed Google Cloud spend.

Sources

• OpenAI — “OpenAI models are coming to Google Cloud” (primary announcement)
• Google Cloud blog — announcement of OpenAI models in Model Garden (primary announcement)
• The Next Web — “OpenAI brings GPT-6 Astra and Codex to Google Cloud, a rival’s platform,” Ana Maria Constantin, 21 October 2026
• Techzine — “OpenAI’s models are finally fully supported on Google Cloud,” Erik van Klinken, 21 October 2026
• Google Cloud — Flexible Savings Plans documentation
• OpenAI — “OpenAI on Oracle Cloud” (primary announcement, June 2026, updated August 2026)

Please follow and like us: