OpenAI's GPT-6 Models on Google Cloud Model Garden: Private Preview Now, General Availability in Weeks

Bottom line: Google Cloud says GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna are rolling out now in private preview to select enterprise customers in Model Garden. SiliconANGLE reports general availability within a few weeks, along with price ratios that Google’s own post does not state.

The blog and the writeup are not describing the same clock. Google Cloud’s October 21 post says GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna are rolling out now in private preview to select enterprise customers in Model Garden. SiliconANGLE the same day says Astra, Sol, and Luna will become generally available in Model Garden within a few weeks. Private preview now is not general availability in a few weeks. One is a limited rollout with a form. The other is a later switch the blog does not date.

We have already watched a ceiling get written as a default. The hour on API Gateway was a maximum, and SSE still defaulted to 15 seconds. We have already watched a sticker that was not the bill. Haiku 5.5 lists at $0.10. Model Garden’s “few weeks” is the same shape of mistake if you paste it over a sentence that says the preview is already rolling.

Private preview is the sentence with a date

The Google post is dated October 21, 2026. The getting-started paragraph names three models, not one. GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna. “Rolling out now.” “Private preview.” “Select enterprise customers.” “Model garden,” lowercased on that line. “With more coming soon.” Then a second sentence: starting today, customers can build with these models through the APIs and connect local Codex CLI and Desktop workflows to the Gemini platform for inference. There is a form for interest.

“Starting today” is October 21, and it is attached to APIs and to Codex CLI and Desktop, for the customers in that preview. It is not attached to every Google Cloud project. “Select” is doing work. A project that cannot see the models on October 27 is not a contradiction of “rolling out now.” A rollout to select accounts is allowed to be invisible to the accounts that were not selected. What you cannot do is read SiliconANGLE’s “within a few weeks” as the start date. That sentence is about general availability. The Google paragraph already started something else.

SiliconANGLE’s piece, updated 20:22 EDT on October 21, says the first LLM OpenAI is bringing to Google Cloud is flagship GPT-6 Astra, then that customers will also receive GPT-6.1 Sol and GPT-6 Luna, then that the three “will become generally available” within a few weeks. “Will become” is future. “Rolling out now” is present. Both can be true if preview and general availability are different gates. They cannot be true if a status page has only one field and you fill it with the future sentence.

shattered.io’s partnership note is the writeup that keeps both gates. Private preview to select enterprise customers, and general availability planned for the coming weeks, with no specific date published by either company as of that piece. That is the reading that does not delete the Google verb. tech-insider’s October 21 FAQ also says general availability is “in the coming weeks,” without a confirmed date as of October 21. Its FAQ on which model is “specifically named” answers GPT-6 Astra, alongside Codex. The Google paragraph names three models. Astra is one of them. A FAQ that says Astra is the model cited as an example is narrower than the sentence it is summarizing. Codex is in the Google paragraph too, as CLI and Desktop workflows, not as a fourth model name.

Coming soon, on the Google post, is a different list. OpenAI models in Gemini for business, and in Google Antigravity. Confidential Compute coming soon. Zero Data Retention options available coming soon. Those are not the Model Garden preview. If a plan says “we will wait for general availability,” it is waiting on a date nobody printed. If a plan says “we will wait for Antigravity,” it is waiting on a product the blog put in the coming-soon sentence, not in the rolling-out-now sentence.

Five times and a hundred times are not on the blog

SiliconANGLE prints a price relationship the Google post does not print. Sol and Luna “cost five and 100 times less than Astra, respectively.” Sol is the five. Luna is the hundred. “Times less” is a slippery English phrase. Five times less than a price can be read as one-fifth, or as a subtraction of five times the price, which is nonsense for a positive sticker. In a pricing sentence next to “scaled-down versions,” the intended reading is almost certainly Sol at one-fifth of Astra and Luna at one-hundredth. The article does not give Astra’s price. A ratio without a base is not an invoice.

No five. No hundred. No “times less.” shattered.io uses “near-Astra intelligence at a lower cost” for GPT-6.1 Sol, which tracks the blog’s adjective, and calls Luna the faster, lower-cost option. Then it has a different sentence: GPT-6 Sol and Luna “launched at roughly half the price of their predecessors.” Half of a predecessor is not one-fifth of Astra, and it is not one-hundredth of Astra, unless the predecessor prices happen to make those ratios land on the same numbers. Nobody in these four pages shows that arithmetic.

Notice the name in that half-price sentence. shattered.io writes “GPT-6 Sol,” not “GPT-6.1 Sol.” The Google post and SiliconANGLE write “GPT-6.1 Sol” for the model entering Model Garden. tech-insider’s table splits them on purpose. GPT-6 Sol and GPT-6 Luna: September 22, 2026, introduced for ChatGPT Work and Codex, rolled out to Plus, Pro, Business, Enterprise, and Edu. GPT-6.1 Sol: an October 2026 update that added local and SSH thread support inside Codex. If that table is right, GPT-6 Sol and GPT-6.1 Sol are not a typo. They are a September product and an October update. SiliconANGLE says Sol and Luna were “launched in late September” and then uses the name GPT-6.1 Sol for the Model Garden pair. Late September matches the table’s September 22 line. The “.1” matches the October row. One sentence is carrying both.

The Google post says GPT-6.1 Sol. A console that only lists GPT-6 Sol is not listing the string the blog printed. A console that lists both is listing the table’s split. Copy the id from the console. Do not copy “Sol” from a headline and assume the minor version.

The half-price sentence and the 5x/100x sentence also compare against different objects. Half of predecessors. Five and a hundred times less than Astra. Predecessors are earlier models. Astra is the flagship in the same announcement. A buyer who writes “Luna is 100x cheaper” in a ticket has chosen SiliconANGLE’s ratio and SiliconANGLE’s base. A buyer who writes “about half” has chosen shattered.io’s other sentence, and that sentence says GPT-6 Sol, not GPT-6.1 Sol. The Google post lets you write “lower cost” and nothing sharper.

”Most advanced” is the headline, not the blog

SiliconANGLE’s headline says OpenAI will make its most advanced AI models available in Google Cloud. The body then calls Astra the flagship and calls Sol and Luna scaled-down versions that trade some output quality for a lower token price. A headline that says “most advanced” and a body that says two of the three are scaled-down are not the same claim. The scaled-down pair can still be part of a launch. They are not, on SiliconANGLE’s own wording, the most advanced model in the set. Astra is the one that sentence reserves for flagship.

The Google post does not say “most advanced.” It says the three names, private preview, and a platform claim beside them: Google Cloud is now the only hyperscale cloud provider offering access to all three leading closed-source LLM families, alongside more than 200 first-party, third-party, and open models. “All three leading closed-source” families is a vendor-count claim. It is not a ranking of Astra against Sol. SiliconANGLE also says Model Garden provides access to more than 200 neural networks. The “more than 200” figure is the part those two pages share. The “most advanced” phrase is not.

tech-insider’s distribution count is a fourth number, and it can sit next to the hyperscaler count without being the same count. Before this week, the piece says, Astra and other recent OpenAI models were already on the OpenAI API, Microsoft Azure, and AWS Bedrock. Google Cloud is then “the fourth distribution channel, and the third major hyperscaler.” Four channels, if you count the API. Three hyperscalers, if you count Azure, AWS, and Google. The Google post’s “only hyperscale cloud provider” offering all three closed-source families is a claim about the set of hyperscalers, not a claim that Google invented the fourth door. Paste them into one cell and the cell will argue with itself.

The writeup says Astra achieved a perfect score on a benchmark that measures finding and exploiting flaws, and that OpenAI says the model causes “fewer misaligned outcomes” than the other frontier models it has evaluated, and that it is configured not to bypass user-defined restrictions. Those are attributions to OpenAI via SiliconANGLE. A security review that needs the score needs the OpenAI source, not a cloud announcement that did not carry it.

Codex, Antigravity, and the product that is not Gemini Agent

Codex is in the Google sentence as something you connect, not as a model id. Local Codex CLI and Desktop workflows, pointed at the Gemini platform for inference. shattered.io treats Codex as the other half of the deal, a multi-step agent, and says local CLI and desktop access is part of what ships with the preview. tech-insider’s October row for GPT-6.1 Sol is about local and SSH threads inside Codex. SSH threads and “connect your local CLI to Gemini for inference” are neighboring claims. They are not automatically the same flag. One is a Codex product update dated, in that table, to October. The other is a cloud announcement dated October 21 that tells you to point the CLI at Google’s platform. A runbook that says “SSH threads turned on because Model Garden went to preview” has joined two sentences that did not join themselves.

Antigravity is where the product name slips. The Google post says OpenAI models will be available in Google Antigravity, coming soon, and in Gemini for business, coming soon. SiliconANGLE says OpenAI plans to make its models available via Google’s Antigravity and Gemini Agent products, and then defines Antigravity as a vibe coding tool that uses agent ensembles, and Gemini Agent as a product that automates business tasks across cloud applications. “Gemini for business” and “Gemini Agent” are not the same string.

“Vibe coding tool” is SiliconANGLE’s description. If your security review is about whether Antigravity is in scope this week, the blog’s answer is coming soon, not private preview. The preview sentence is Model Garden, APIs, and Codex CLI and Desktop. Antigravity is in the next paragraph.

Confidential Compute is also coming soon, on both the Google post and SiliconANGLE. SiliconANGLE says the companies plan to follow the launch by bringing the models to Confidential Compute. The blog says OpenAI models run on dedicated Google Cloud infrastructure with Confidential Compute coming soon, and that queries are routed through isolated gateways. “Dedicated infrastructure” now, and Confidential Compute coming soon, can both be in one bullet. A design doc that checks the Confidential Compute box because the preview started has checked the wrong half of the bullet.

What to write in the tracker

Model ids, from the Google post: GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna. Status, from that post: private preview, select enterprise customers, rolling out now as of October 21, interest form, APIs and local Codex CLI and Desktop in the “starting today” sentence. General availability: “within a few weeks” or “in the coming weeks,” no day, from SiliconANGLE, shattered.io, and tech-insider.

Prices: do not copy 5x and 100x into a budget sheet unless SiliconANGLE is your source and you have Astra’s rate from somewhere else. Do not copy “roughly half” onto GPT-6.1 Sol. The half sentence says GPT-6 Sol and predecessors. The blog says lower cost.

Names to keep apart: Gemini for business versus Gemini Agent. Antigravity coming soon versus Model Garden preview now. Fourth distribution channel versus only hyperscaler with all three closed-source families. Astra as flagship in a writeup versus “most advanced” in a headline that also covers two scaled-down models.

The preview is the thing with a start. The weeks are the thing without a day. A status field that holds only one of those sentences is unfinished. Flint’s 150 and NVIDIA’s 220 were two integers that did not share a row. These are two clocks, and a price ratio the announcement did not print.

What this means for you

If your organization wants early access, request it through Google’s interest form now; otherwise plan around general availability with no fixed date. Copy model IDs and prices from the console, not from news summaries, when you update budgets and configs.