If you have a Google Cloud commit you are behind on, OpenAI's models just became a way to spend it. The models are GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, and they sit in private preview with no published price, no GA date, no listed regions, and zero data retention still marked "coming soon." Google's own savings-plan terms also give third-party models no discount. The practical answer for most enterprises: use the preview to run evaluations and to sharpen your Azure or Bedrock renewal, and keep regulated production traffic where its controls are already in writing.
Google Cloud announced the launch on October 21, 2026, in a post by Michael Gerstenhaber, Google Cloud's VP of product management for Gemini. With it, OpenAI's frontier models are now sold through all three hyperscalers. What follows covers what the announcement commits to, what it leaves blank, and how to use it in the next budget cycle.
What Google Actually Shipped on October 21
Google shipped a private preview for selected enterprise customers, a billing path, and a list of controls, several of them future tense. According to Google's announcement, three OpenAI models are entering Model Garden: GPT-6 Astra for coding, computer use and research work, GPT-6.1 Sol as a cheaper near-Astra tier for agentic coding, and GPT-6 Luna for high-volume everyday tasks. Access starts with an interest form, and OpenAI models in Gemini for business and in Google Antigravity are listed as coming later.
The commercial terms are the part most buyers will care about. The models are transacted through Google Cloud Marketplace, appear on a single Google Cloud invoice, draw down existing cloud commitments and count toward Flexible Savings Plans. Codex users can point the Codex CLI and Desktop app at Google's platform for inference starting on launch day, so a developer team can test against Google billing this week.
Google says prompts, context and outputs stay inside the customer's data perimeter and are never used to train foundation models, that the models run on dedicated Google Cloud infrastructure behind isolated gateways, and that every interaction runs under an authenticated identity with VPC Service Controls and unified audit logging. Zero Data Retention options and Confidential Compute are both labeled "coming soon." Techzine's report adds that those two features still need additional work.
On timing, The Next Web reports that OpenAI plans general availability "in the coming weeks." Google's own post gives no GA date. Neither source publishes a per-token price, and Google's post names no regions.
Why the Commit Drawdown Matters More Than the Models
For a company running short on its Google commit, the drawdown is the whole story. OpenAI usage that fills a shortfall costs nothing extra, because that money gets paid either way. Google's Flexible Savings Plans documentation (last updated October 20, 2026) is explicit that "Unused commitment does NOT carry over to the next month" and that you pay the full monthly commitment amount, including the unused portion. If your engineers already want GPT-6 Astra and your Google commitment is trailing, routing that usage through Marketplace converts a write-off into model access.
The same page also limits the upside. A one-year FSP gives a 10% discount and a three-year FSP gives 20%, but only on the listed Gemini Enterprise SKUs, and the documentation states that "Third-party products don't get any discounts from the FSP." OpenAI models count toward the commitment, so they burn it down at whatever rate Google charges. They do not earn the FSP discount Gemini earns. Usage above the commitment is billed at on-demand or negotiated rates with no FSP discount on the overage.
That rate is the open question. On OpenAI's own API pricing page, GPT-6 Astra lists at $10 per million input tokens and $50 per million output tokens, GPT-6.1 Sol at $2 and $10, and GPT-6 Luna at $0.10 and $0.50, with Batch and Flex at half those rates. OpenAI also charges 10% more on regional data-residency endpoints for models released on or after March 5, 2026. Google has published none of its own prices for these models, so you cannot yet tell whether Marketplace matches OpenAI's list, adds a residency uplift, or offers Batch at all.
AWS answered that question in writing. When GPT-5.6 Sol, Terra and Luna went generally available on Amazon Bedrock on July 13, 2026, the announcement said "Pricing matches OpenAI first-party rates and usage counts toward your AWS commitments." Ask Google for the same sentence before you plan a budget around it. If you are weighing reserved capacity instead of on-demand, our provisioned throughput break-even analysis covers the utilization a monthly term needs.
Can You Run Regulated Workloads on It Yet?
Not on the record Google has published so far. The announcement lists zero data retention and Confidential Compute as future features, and Google's Vertex AI security controls page (last updated October 22, 2026) carries a blunt caveat: "Security controls are not supported for preview models." That page covers data residency at rest, customer-managed encryption keys, VPC Service Controls and Access Transparency. The launch post claims VPC Service Controls protection for the OpenAI models. The general documentation says preview models fall outside those controls. Until Google reconciles the two in writing, treat the narrower statement as binding.
Zero data retention (ZDR) is a commitment that the provider stores no prompts or outputs beyond the request, including for abuse monitoring. On Google's platform it is partly a configuration exercise. Google's data governance page says that "To achieve zero data retention, customers must take specific actions," such as leaving request-response logging off and setting store = false on Interactions API calls, and it notes that Grounding with Google Search stores data with no way to disable it. The same security controls page says prompt logging for partner and open models "is turned off by default" and follows the platform's ZDR policy. The OpenAI launch post, though, lists ZDR as "coming soon" for these models specifically. For now, OpenAI's models on Google sit outside the partner-model ZDR path your security team may already have approved for Claude.
Compare that with the controls on your existing OpenAI route. Microsoft's data privacy page for models sold by Azure documents how managed customers apply for modified abuse monitoring, which turns off the human-review data store, and how to verify it with the ContentLogging attribute. It also warns that "Azure Preview features, including Models sold by Azure in preview, may employ different privacy practices." That warning is a fair description of where Google's OpenAI offering sits today. Our LLM data residency guide walks through which providers keep processing in region and how abuse monitoring changes that answer.
What the Bedrock Rollout Says About the Wait
The closest precedent took roughly ten weeks from preview to GA. OpenAI's models reached Amazon Bedrock in limited preview on April 28, 2026, after the end of Azure's exclusivity that we covered in OpenAI's move to multi-cloud. The first generally available OpenAI models on Bedrock, the GPT-5.6 family, arrived in July. Even then, AWS launched Sol in two US regions and Terra and Luna in three. The same Amazon page says GPT-6 Astra became generally available on Bedrock on September 8, 2026.
If Google follows the same arc, GA lands around the turn of the year, and in-region inference outside the US could come later still. If OpenAI's "coming weeks" holds instead, regions and prices will still arrive in the GA notice rather than before it. Either way, a budget line written now is a guess.
The strongest argument for moving early belongs to Google's side. Google says it is now the only hyperscaler offering Gemini, Claude and GPT through one platform, and The Next Web cites Reuters reporting that Anthropic earned 47% of its 2025 revenue through Amazon's and Google's clouds, according to its IPO filing. Frontier labs clearly want hyperscaler distribution. A single platform with all three model families makes it cheaper to switch models per task, and that has real value if your evaluations show Gemini 4 Argon winning some workloads and Astra winning others. Our look at Gemini 4 Argon's cost per task is a starting point for that comparison.
That argument holds up for evaluation work. It is weaker for production, where the deciding facts are price, region and retention terms, and Google has published none of the three for these models.
Who Gains Leverage From This
Your Azure and AWS account teams just lost their monopoly on enterprise OpenAI access, and that is negotiable. OpenAI still calls Microsoft its primary cloud partner, but a customer with a credible plan to move GPT-6 Astra usage to Google has a concrete alternative to cite when its Azure MACC or AWS EDP comes up for renewal. Our piece on how DMA duties land midway through a three-year commit explains why the renewal is the only point where that leverage converts into terms.
The leverage is only as good as your evidence. An account team will discount a vague threat to "look at Google." A side-by-side evaluation of the same workload on GPT-6 Astra via Google Vertex AI and via Microsoft Foundry or Amazon Bedrock, with latency, error rate and the price each vendor quoted, is harder to dismiss. Moor Insights & Strategy analyst Mike Leone, quoted via Axios, framed the pressure from Google's side: "Gemini now has to win those customers with OpenAI sitting right next to it." Google's own account team now faces the same comparison, which helps you on the Gemini side of the table too.
What to Do Before GA
This Week:
- Ask finance for your Google Cloud commit burn rate and projected shortfall at term end. If you are on track, the drawdown changes nothing for you and you can stop here.
- If you are short, submit the interest form and ask your Google account team three written questions: the per-token price for each OpenAI model on Marketplace, the regions at GA, and the date ZDR and Confidential Compute become available.
- Point one team's Codex CLI at Google's platform for a week and compare the bill line with what the same work costs on your current route.
This Month:
- Have your security lead reconcile Google's "Security controls are not supported for preview models" with the launch post's VPC Service Controls claim, and get the answer in writing before any non-public data touches the preview.
- Run your existing OpenAI evaluation set on the Google preview and your current provider, and record latency, error rate and output parity.
- Check whether your Google contract treats Marketplace third-party spend as fully eligible for commit drawdown. The FSP page covers savings plans, and your negotiated commit terms may differ.
Before Your Next Azure or AWS Renewal:
- Bring the side-by-side results and Google's written price to the table, and ask for matched pricing or better commit terms on OpenAI usage.
- Keep regulated production on the route whose retention terms are already signed until Google publishes ZDR for these models. Our questions for AI vendor security review cover what to get on paper first.
The Bottom Line
OpenAI's route onto AWS ran in a clear order: announcement and commit billing in April, then published price parity and named regions at GA in July. Google's offering is at the first step. The one hard financial term it publishes, FSP eligibility without the FSP discount, mostly helps customers who were going to waste part of their commit anyway.
If that is you, start the evaluation now and put the shortfall to use. If it isn't, use the launch in your renewal and wait for Google to publish the price and the ZDR date.
Continue Reading
- Provisioned Throughput Pricing: Monthly Terms Need 99% Utilization
- LLM Data Residency: Which Providers Actually Keep Data In Region
- OpenAI Goes Multi-Cloud: AWS and Google Win Enterprise Access Through 2032
- Model Deprecation: Bedrock Keeps Claude 4 Months Past Anthropic
- OpenAI Revenue Runs at $50B Once Partner Sales Come Out
- Fable 5 Needs Retention On. ZDR Was Never Zero.