OpenAI models head to Google Cloud and Antigravity

By VK

Published 22 Oct 2026, 10:40 UTC

Reporting from OpenAI, Ars Technica

OpenAI's Astra, Sol and Luna enter private preview in Google Cloud's model garden, billed through Google, with Antigravity support to follow.

What it means for founders

Committed cloud spend just got more flexible. If you carry a Google Cloud spending commitment, you will be able to trial GPT models once access opens, without signing a new vendor contract or paying a separate bill; for now only select enterprise customers are in the preview. Neither company has said whether startup credits apply, so ask your Google account team before you plan around them.
Multi-model routing is becoming the norm. With GPT, Claude and Gemini all on one Google platform, comparing models task by task needs fewer contracts and less integration work. Keep prompts and evaluations portable so that a switch means changing a setting rather than rewriting code.
Compliance answers are moving to the cloud layer. Regulated buyers may accept GPT more readily when it runs inside their own Google perimeter with audit logs and VPC Service Controls. If enterprise deals stall on data handling questions, this gives you a new answer, though the retention controls are not finished.
Watch the general availability date. The signals to track are the price list at launch, any response from Microsoft to OpenAI selling through a rival cloud, and when Antigravity's model picker lists GPT.

The story

OpenAI and Google Cloud said on Wednesday, October 21, that Google Cloud's model garden will start selling OpenAI's frontier models. The garden is Google's catalogue of more than 200 first-party, partner and open models. In its announcement, OpenAI said Codex workflows are part of the arrangement and that Microsoft remains its primary cloud partner. Google's post says a limited group of enterprise customers can use them today in a private preview, and OpenAI expects general availability within weeks.

What Google Cloud customers get

Google lists three models in the first batch: Astra for demanding coding, computer use and research work, GPT-6.1 Sol, which Google pitches as nearly as capable as Astra for less money, and Luna for quick, everyday jobs in high volume. Google says the OpenAI models are served from hardware set aside for them inside its cloud, and that customer prompts and outputs stay within the customer's own boundary and do not feed model training. Zero data retention and confidential computing options are still to come.

Billing is the bigger shift. Customers buy through Google Cloud Marketplace, so a company that has already promised Google a set amount of spending can put part of it toward GPT models, and the charges show up on its normal Google bill. Customers can also point the Codex command line tool or desktop app at Google's platform for inference, according to Google, which did not say whether that option widens with general availability.

Antigravity and the race for agent platforms

Both companies say OpenAI models will reach Antigravity, Google's agentic coding environment, at a later date. Ars Technica reported that the tool already offers Gemini and Claude, and that Google has not said which OpenAI models it will add. Google now describes itself as the only hyperscale cloud offering all three leading closed model families. Google's post names one customer so far: Box, which says it plans to offer its own users OpenAI models through Google Cloud.

What we don't know yet

What OpenAI models will cost on Google Cloud, and whether prices will match OpenAI's own API.
When access widens beyond the private preview, other than OpenAI's plan for general availability in the coming weeks.
Which GPT models Antigravity will offer, and on which plans.
When the zero data retention and confidential computing options will ship.