On October 21, Google Cloud put OpenAI’s GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna into its model garden, and the first sentence of Google’s own announcement tells you how far along that is: the models are “rolling out now in private preview to select enterprise customers.” There is an interest form. There is no published price, no list of regions, and, on launch day, no GPT-6 entry in Google’s partner-model documentation, according to a read-through of the vendor pages by Digital Applied.
The headlines called it the moment OpenAI finished its escape from Microsoft exclusivity. That was the story in May, when the multicloud deal was signed. What shipped this week is narrower and, for anyone who signs cloud contracts, more useful to understand precisely: OpenAI on Google Cloud is a billing feature before it is a model feature. The model is the same one you can already call elsewhere. What changed is which invoice it lands on.
The model was already on the other two clouds
GPT-6 Astra has been generally available on Amazon Bedrock since September 8, and Microsoft made it generally available to all Foundry customers in September, with Standard and Provisioned Throughput options in Global and US Data Zone geographies and a published list price of $10 per million input tokens and $50 per million output on Standard Global. Google is the third hyperscaler to carry it, and the last.
So the capability question was settled weeks ago. If you wanted Astra for its coding and computer-use strength (and the cost-per-task math we ran on it in September still holds), you could already have it under an AWS or Azure agreement. Google’s pitch is different, and it is stated plainly in the announcement: because the models are sold 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.”
That sentence is the deal. A company that signed a multi-year spend commitment with Google, and whose engineers would rather call GPT-6.1 Sol than Gemini for some workloads, no longer has to choose between using the model it wants and burning down the commitment it already owes. That is a real problem in real procurement offices, and Google solved it. It just isn’t a model launch.
Google also claims it is “now the only hyperscale cloud provider offering access to all three leading closed source LLMs.” Digital Applied’s note on that line is fair: in a private preview with an interest form, “access” is carrying most of the weight.
Drawdown is not a discount
The most common misreading of “counts toward Flexible Savings Plans” will be that OpenAI tokens get cheaper on Google. Nothing in the announcement says that. Google’s own Flexible Savings Plans documentation describes products that draw down an FSP commitment without receiving any incremental discount from it, and names Vertex AI Provisioned Throughput SKUs as examples. Whether OpenAI usage on Google lands on the discounted side of that line or the drawdown-only side is exactly the kind of detail a preview announcement leaves out.
Which brings the missing rate card back into focus. Azure publishes a list price for Astra. Google, as of this week, publishes nothing. Until it does, the comparison a buyer can make is between a known price on one cloud and a commitment-accounting benefit on another, and those aren’t the same unit. The question to ask your Google account team is not “can we use OpenAI models?” It is “what is the per-token rate for each of the three models in the regions we run, and does that usage earn the FSP discount or only consume the commitment?” Get both answers in writing before anyone moves a workload.
There is a second number to keep in view. The same model sold through three channels will not necessarily cost the same through all three, and the gap is a channel decision, not a model decision. Price the channel. A team that assumes “GPT-6 Astra costs $10/$50” because that is the Azure list price may find a different figure on its Google invoice in either direction.
The data controls are listed as coming soon
For regulated work, the paragraph that matters is the one about data handling, and it reads in two tenses. In the present tense, Google says customer prompts, context and outputs “remain entirely within your designated data perimeter and are never used to train foundation models,” with VPC Service Controls and unified audit logging around the gateway. In the future tense, the same sentence continues: “with Zero Data Retention (ZDR) options available coming soon.” A separate line says the models run on dedicated Google Cloud infrastructure “with Confidential Compute coming soon.”
“Never used to train” and “zero data retention” are different promises. The first says your data won’t improve the model. The second says your data isn’t kept at all beyond the request. If your legal team approved an AI vendor on the strength of a ZDR clause (the same clause that became a sticking point when Anthropic made 30-day logging mandatory on its top models), the Google route doesn’t yet offer the clause they signed off on. Neither does it say when it will.
There is also a quieter change for developer tooling. Google says that “starting today,” customers can “connect their local Codex CLI and Desktop workflows directly to our Gemini platform for inference.” That means a developer’s coding agent can be pointed at a new inference endpoint with a configuration change. For a security team, that is a new place source code goes. Before anyone flips it, confirm that the Google endpoint’s retention and residency terms match whatever your current Codex route was approved under. A preview with ZDR “coming soon” may not.
Who should do anything this month
Three kinds of teams have a reason to fill in the interest form now.
The first is a team with an underused Google commitment and a workload that already runs on GPT-6.1 Sol or Luna through OpenAI directly. Moving that spend onto the Google invoice is pure procurement upside, provided the per-token rate comes back close to what you pay today.
The second is a team that standardized on Vertex AI for governance (IAM, VPC Service Controls, audit logging) and has been running a separate, less governed path to OpenAI on the side. Folding that path under the controls you already operate is the better architecture, even before price is known.
The third is a team evaluating whether a mixed-model routing setup can live entirely inside one cloud. With Gemini, Claude and now OpenAI models in one garden, that is closer to possible on Google than anywhere else, once general availability arrives. OpenAI’s own announcement puts that “in the coming weeks.”
Everyone else, and especially anyone whose approval depended on zero data retention, should wait for three things: a published rate card, a named region list, and ZDR moving from “coming soon” to a contract term. None of those is exotic. All three are missing today. When they land, this becomes a real choice about where to buy the same model. Until then it is a preview of a billing arrangement, and a good one to understand before your next commit renewal, not a reason to move production.