What Google shipped, and who can get it
Google Cloud’s announcement on 21 October 2026, written by Michael Gerstenhaber, VP Product Management, Gemini, puts three OpenAI models into Google Cloud’s model garden: GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, “rolling out now in private preview to select enterprise customers”.
Two surfaces are involved, and Google uses both in the same paragraph. The model garden is the catalogue where a customer selects a model. Google Cloud Marketplace is where the purchase is transacted and billed. OpenAI’s models are transacted directly through Google Cloud Marketplace, so usage lands on a Google Cloud invoice rather than a separate OpenAI one.
Access runs through an express-interest form rather than a self-serve sign-up. Google has published no general availability date and no price list for these models on Google Cloud. Customers taken into the preview can call the models through Google’s APIs and connect local Codex CLI and Desktop workflows to Google’s Gemini platform for inference, which Google describes as available to those customers “starting today”.
Independent coverage reads looser than Google’s own wording. Ars Technica’s 21 October report says the models “will appear in Google’s ‘Model Garden’ in the coming weeks”, while noting in the same piece that Google says all three are “rolling out now in private preview for select enterprise customers”. SiliconANGLE wrote that the models “will become generally available in Google’s Model Garden within a few weeks”. Google has said neither of those things. Until it publishes a date, the private preview and the express-interest form are the current state.
The Claude leg of the catalogue is already generally available
Two days after the OpenAI post, Google published a second announcement on Claude: general availability of an Anthropic-managed deployment option for Claude models on Google Cloud, joining the Google-managed option that was already there. Both use Google Cloud IAM and Marketplace billing, and both let a customer consolidate spend under existing Google Cloud billing with eligible commitment drawdown. The models named are Claude Fable 5.1, Sonnet 5.5, Haiku 5.5 and Opus 5.5.
That asymmetry carries the story. The Claude side of the catalogue is generally available and commitment-eligible. The OpenAI side is a preview for select customers, so Google’s claim to offer all three closed families rests on a combination in which its own GPT access is the gated part.
One invoice, one commitment: the procurement change
Google’s sentence on the mechanics reads:
“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.”
Commitment drawdown means committed Google Cloud spend can be consumed by OpenAI model usage; SiliconANGLE describes that as easing procurement by removing the need for a separate AI budget. Flexible Savings Plans are Google Cloud’s spend-based committed use discount. Google’s billing documentation describes a one-year commitment at a 10% discount or a three-year commitment at 20%, with monthly entitlement windows, no rollover and no cancellation. Google says the FSP counting works “across model providers”.
That same document supports the general mechanism with a qualification the blog leaves out. Spend on participating third-party foundation models available through Vertex AI and Cloud Marketplace draws down the enterprise commitment “subject to the standard Cloud Marketplace drawdown rules and contract caps defined in your enterprise agreement”, and “third-party products don’t get any discounts from the FSP”. The supported reading is therefore drawdown and counting, not a discount on OpenAI model usage. The documentation never names OpenAI, so the 21 October blog remains the only source tying the mechanism to these specific models.
One question stays open: whether drawdown and FSP counting apply while the models are in private preview or only at wider availability. Neither the blog nor the billing document says, and Google does not say elsewhere. The commitment wording also carries a source limit. SiliconANGLE attributes the mechanic to OpenAI rather than to Google: “OpenAI says that Google Cloud customers can use those commitments to pay for access to its models.” OpenAI’s own announcement post could not be retrieved, returning HTTP 403, so its exact wording survives only at second hand, through 9to5Google’s quotation.
What changes here is procurement: a rival lab’s models paid for through Google’s invoice and drawn against an existing Google Cloud commitment. A per-agent spend limit governs how much one agent may consume; it does not decide which invoice the model usage lands on.
Customer testimony in the announcement comes from Box. Yashodha Bhavnani, VP of AI Products, calls the move “a step toward operating more efficiently”. Box is the only named customer in Google’s post, and the quote was supplied by Google: on-record customer testimony rather than independent evidence.
What is still only “coming soon”
Four items sit beyond the preview, and none of them is shipped:
- Zero Data Retention options, which Google describes as “coming soon”.
- Confidential Compute for these models, “coming soon”.
- Availability of OpenAI models in Google Cloud’s Gemini for business.
- Availability in Google Antigravity, alongside “more coming soon” on wider availability.
The platform assurances are a different category. Google says customer prompts, context and outputs remain within the customer’s designated data perimeter and are not used to train foundation models; that queries route through hardened, isolated gateways; and that every agent and model interaction runs under an authenticated identity with VPC Service Controls protection and unified audit logging. Those are Google’s present-tense statements about its own platform. No separate Google Cloud documentation was found for them, and none of it has been independently audited.
Google’s “only hyperscaler with all three” claim, in Google’s own words
The sentence at the centre of the announcement reads:
“Google Cloud is now the only hyperscale cloud provider offering access to all three leading closed source LLMs.”
Google’s own share-text variant of the same sentence names them: Gemini, Claude and GPT. “Leading” is never defined, and the boundary of the category is Google’s to draw.
The claim is checkable against the two rivals’ catalogues. Microsoft’s Foundry lists the GPT-6 family (Astra, Sol, Luna and 6.1 Sol) alongside a long list of Anthropic Claude models, transacted through Azure Marketplace, and no Gemini model appears in either Foundry catalogue. Amazon Bedrock lists Claude and OpenAI’s GPT-6 models, with Google represented only by the open Gemma models. On the narrow axis Google states, catalogue access to all three closed-source families, neither Azure nor AWS currently offers Gemini.
Three qualifications belong with that finding. Google’s own GPT access is the select-customer preview, while Azure and AWS list GPT-6 as available in their catalogues. Amazon started hosting OpenAI models in April 2026, after the Microsoft exclusivity ended, as CNBC reported, so Google is not the first cloud platform to host OpenAI’s frontier models. And nobody independent has tested the claim: Ars Technica repeats it with attribution to Google, and no other outlet reviewed for this article went further.
As of 24 October 2026, Google’s assertion is not contradicted by the public catalogues and is not third-party verified.
Google keeps the workload when a rival’s model does the work
Reselling a rival lab’s frontier models is a billing and procurement move at least as much as an openness move. The customer gets one invoice, one commitment to draw against and one set of Marketplace billing rules. Google keeps the workload even when a competitor’s model does the work, and the FSP terms are non-cancellable with no rollover.
The evidence for that reading is mechanical. No analyst commentary on consolidating multi-model spend into a single cloud commitment was found for this announcement, and no procurement practitioner has published an assessment of it. What the documents show is the structure: third-party usage can consume an enterprise commitment without earning the FSP discount, drawdown runs under Marketplace rules and contract caps, and the plan cannot be cancelled mid-term. Google Cloud CEO Thomas Kurian framed the announcement as selection: “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”, as SiliconANGLE reported.
What to pin down before committing spend
Four questions decide what a preview customer is agreeing to:
- Whether usage during the private preview is eligible for commitment drawdown and FSP counting, or whether that starts at wider availability.
- When the models reach general availability on Google Cloud, and at what price. No price list exists.
- When Zero Data Retention and Confidential Compute arrive for these models. Google says “coming soon” with no date.
- How much of an enterprise commitment third-party model spend can consume, and which Marketplace drawdown caps and contract limits apply.
Google has published answers to none of them, and the express-interest form remains the documented way in.
