Daily · Article 03 · 5 min read
- OpenAI
OpenAI models enter private preview for selected Google Cloud enterprises
Previously uncovered October 21 announcements and overnight changes · rechecked at 06:00 JST October 22
Summary
OpenAI and Google Cloud announced on October 21 and October 22 respectively that OpenAI models are coming to Google Cloud's Model Garden. For now, only selected enterprise customers can use GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, in private preview.
Say your company runs its services on Google Cloud and wants to try GPT models. A route is opening to pick OpenAI models inside Google Cloud and have the usage show up on your Google Cloud bill. For now, though, it's a private preview for a limited set of companies.
Key facts
OpenAI on October 21 and Google Cloud in a post dated October 22 announced that OpenAI models are coming to Google Cloud's Model Garden, the part of Google Cloud where you pick and use AI models from several companies. According to Google, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are rolling out in private preview to selected enterprise customers, and interested companies can register through Google's interest form. OpenAI announcement · Google Cloud announcement
| Item | Status | Who said so |
|---|---|---|
| GPT-6 Astra · GPT-6.1 Sol · GPT-6 Luna | Private preview for selected enterprise customers, with more models coming soon | |
| Building through the APIs, and connecting local workflows in Codex, OpenAI's coding agent (CLI and desktop app), to Google's Gemini platform to call the models | Google says “customers” can do this from the announcement date; it doesn't say whether this is limited to the preview customers | |
| OpenAI models in Gemini for business and in Antigravity, Google's agentic development tool | Coming soon | |
| Zero Data Retention (ZDR) options, which keep no data, and Confidential Compute | Coming soon | |
| General availability | In the coming weeks | OpenAI (Google's post gives no date) |
1. The path a company app's request takes
Google Cloud is where companies run their servers, data and applications. With this announcement, a company app running there can call OpenAI models. Google says the models run on dedicated Google Cloud infrastructure, and that every call is made under an authenticated identity and recorded in audit logs, the records that let you trace later who did what and when. Google Cloud announcement
The cat asks Q&A
The researcherIf I pick GPT inside Google Cloud, is Gemini the one actually answering?
Pocket NotebookNo, the OpenAI model computes the answer. Google Cloud sits in front of it and handles who can call it, what gets logged and what gets billed.
- The researcher
Our company is a Google Cloud customer. Can we start using it today?
Pocket NotebookOnly selected companies have preview access at the moment. If you're not one of them, register interest and wait for general availability.
2. Billing and commitment discounts
According to Google, OpenAI models are sold through Google Cloud Marketplace, so usage appears on a single Google Cloud invoice and draws down existing cloud commitments. A cloud commitment is a contract in which a company agrees to spend at least a set amount with a cloud provider over a period. The usage also counts toward Flexible Savings Plans (FSP), which Google presents as Gemini Enterprise commitment plans: commit to a monthly spend and token costs drop 10% on a one-year commitment or 20% on a three-year one. Google Cloud announcement About FSP
Google's post doesn't list per-model prices. What you actually pay, and which commitment it counts against, depends on each company's contract.
The cat asks Q&A
- The researcher
If we already have a commitment in place, do we not pay separately for GPT?
Pocket NotebookYou still pay for the usage. Instead of a separate bill, it comes out of the amount you've already committed, so heavy use eats into that commitment faster.
3. Where does your input data stay?
Google says customer prompts, context and outputs stay within the data perimeter the customer designates and are never used to train foundation models. Calls go through isolated gateways and stay inside Google Cloud, and third-party model providers have no unauthorized access to the customer's cloud environment. The ZDR options, which keep no data at all, and Confidential Compute, which protects data while it's being processed, are still marked "coming soon." Google Cloud announcement
The cat asks Q&A
- The researcher
If it isn't used for training, does that mean it isn't stored either?
Pocket NotebookThat's a separate promise. The no-storage ZDR option isn't out yet, so for how long data is kept and what gets logged today, look in your contract and the product terms.
4. Which of the three models?
Google describes the three models this way.
| Model | Google's description |
|---|---|
| GPT-6 Astra | For demanding work such as coding, computer use, professional tasks and scientific research |
| GPT-6.1 Sol | Close to Astra at a lower cost, for agentic coding, computer use and professional work |
| GPT-6 Luna | A fast, low-cost model for everyday work at scale |
Giving them the same job tells you more than the descriptions do. For a fictional support-ticket routing app, first decide which categories count as correct and when a person should review the result, then compare format, accuracy, response time and cost on the same inputs. The AnyLanguageModel article covers swapping between models inside an app.
The cat asks Q&A
- The researcher
So why not just pick Astra, the most capable one?
Pocket NotebookIt depends on the job. Going by Google's descriptions, Luna fits high-volume everyday work and Astra fits hard coding or research. Run the same inputs from your own app and decide on accuracy, speed and cost.
To close
What to remember
Companies that run their systems on Google Cloud now have a way to use OpenAI models on the same bill and under the same access controls. This isn't a change to ChatGPT subscriptions, and for now only a limited set of companies can try it in preview. If your team needs a guarantee that no data is kept, it's safer to wait until the ZDR option ships.
Sources
- OpenAI announcementopenai.com
- Google Cloud announcementcloud.google.com
- About FSPcloud.google.com