On October 21, 2026, TestingCatalog reported that Google Cloud was adding three OpenAI models to Model Garden in private preview: GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. The newsroom flagged the story because that third-party report was the only source supplied. The main question was whether Google or OpenAI had officially confirmed the rollout, and on what terms.
This matters for builders because of what it would complete. Gemini is native to Google Cloud and Claude is already on Vertex AI, so adding OpenAI's frontier models would let one cloud contract, one invoice and one set of IAM and VPC controls cover all three major closed model families. Many enterprises choose an AI platform for procurement and compliance reasons, not just benchmarks. Google is the last of the three big US hyperscalers to offer these models.
What we checked
- Can Google Cloud or OpenAI publicly confirm the private preview, eligible customers, model names, billing terms, and data-governance controls? partly confirmed. Google Cloud's official blog (October 21, 2026, by VP Michael Gerstenhaber) confirms a private preview for 'select enterprise customers' through an interest form, names GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, and describes Marketplace billing and the governance controls. It does not say who is eligible, and it gives no pricing, regions or GA date. OpenAI's post exists (TNW quotes Greg Brockman), but it returned 403 to us, so we could not read its terms directly. [1][4]
- Do OpenAI models count toward Google Cloud Flexible Savings Plans discounts, as 'unified billing' might suggest? partly confirmed. The models count toward FSP commitments but get no discount. Google's FSP docs (updated 2026-10-20) say 'Third-party products don't get any discounts from the FSP.' Third-party spend draws down commitments, and Private Offer pricing is the exception. [3][1]
- Are the GPT-6 models listed in public Vertex AI / Agent Platform documentation yet? refuted. No. When we checked, the partner-models documentation listed Claude, Grok, Mistral, Meta and Llama models. The only OpenAI entries were the open-weight gpt-oss-120b and gpt-oss-20b, with no GPT-6 entry. [5]
What we found
- Google officially confirmed the launch. A Google Cloud blog post dated October 21, 2026, by Michael Gerstenhaber (VP, Product Management, Gemini), says the models are 'rolling out now in private preview to select enterprise customers in Google Cloud's model garden.' It names GPT-6 Astra (coding, computer use, professional and scientific work), GPT-6.1 Sol ('near-Astra intelligence at lower cost') and GPT-6 Luna (fast, cost-efficient everyday work). Access is through an interest form at cloud.google.com/resources/interest/openai-models. The post does not say what makes a customer eligible. [1][2]
- The billing claims are accurate, but the savings are smaller than they sound. Google says usage is 'transacted directly through Google Cloud Marketplace,' appears on one invoice, draws down existing commitments and 'counts toward Flexible Savings Plans.' Google's own FSP documentation, updated 2026-10-20, says 'Third-party products don't get any discounts from the FSP.' So OpenAI spend helps you meet a commitment but gets no FSP discount, unless you negotiate a Private Offer. [1][3]
- Two key governance features are promised but not yet available. Google says prompts, context and outputs stay within the customer's data perimeter and are 'never used to train foundation models.' It also says traffic runs through 'hardened, isolated gateways' with VPC Service Controls and unified audit logging. But Zero Data Retention options and Confidential Compute are both labeled 'coming soon,' with no dates. TNW notes that the data-handling assurances are Google's own claims and have not been independently verified. [1][4]
- The models are not yet in public documentation. When we checked, Google's partner-models docs listed Anthropic, xAI Grok, Mistral, Meta and Llama models. The only OpenAI entries were the open-weight gpt-oss-120b and gpt-oss-20b. Google's announcement gives no per-token pricing, no regions and no GA date. [5][1]
- OpenAI's side has been reported, but we could not read its post. TNW reports that OpenAI's announcement quotes president Greg Brockman, plans general availability 'in the coming weeks,' and says 'Microsoft remains OpenAI's primary cloud partner.' Search snippets of OpenAI's post describe a limited preview for development and internal testing through the Responses API and Codex. They also say a Model Garden listing will come at GA. OpenAI's page returned HTTP 403 to us, so we could not check those scope limits ourselves. [4]
- Codex and other Google tools are part of the plan. Google says preview customers can 'connect their local Codex CLI and Desktop workflows directly to our Gemini platform for inference.' OpenAI models are 'coming soon' to Gemini for business and Google Antigravity. Box is the only named customer: its VP of AI Products says Box will pair OpenAI models with Google Cloud's security and compliance controls. [1]
- Google arrived after AWS and Azure. Digital Applied, citing vendor posts we did not open ourselves, reports that GPT-6 Astra went GA on Microsoft Foundry on September 3, 2026, and on Amazon Bedrock on September 8. It also reports that AWS documentation offers zero data retention on request. TNW says AWS has hosted OpenAI models since April, after Microsoft's exclusive license ended. [6][4]
What it means
The launch is real and official, but it is not ready for production planning. If your organization is committed to Google Cloud and wants GPT-6-class models under existing IAM, VPC-SC and audit tooling, submit the interest form and start evaluations. Plan on the preview being restricted: OpenAI's post reportedly limits it to development and internal testing, and there is no published rate card, region list or SLA. Don't count on zero data retention or Confidential Compute until Google ships them. Regulated workloads that need those today are better served by Bedrock or Foundry, which already list these models as GA, as long as their data terms fit your requirements.
For FinOps, be precise. The pitch is about where the money goes, not about lower prices. OpenAI spend draws down your Google commitments and FSPs, but Google's own FSP docs say third-party models get no FSP discount. Any price advantage would have to come from a negotiated Private Offer. Architecturally, the real benefit is being able to switch models: if your code abstracts over Gemini, Claude and GPT behind one gateway, you can route by cost and latency within a single cloud. Build that abstraction now and wait for GA, pricing and docs before moving traffic over.
Still open
- What are the per-token prices on Google Cloud, and do they include a premium over OpenAI's direct API rates?
- Which regions and data-residency options will be offered, and when?
- Who is eligible for the private preview, and what does 'select enterprise customers' mean in practice?
- When will zero data retention and Confidential Compute ship, and on what terms?