OpenAI Comes to Google Cloud: Its Models Are Now Sold in Model Garden
Direct answer
OpenAI announced on October 21, 2026 that its frontier models and Codex agentic workflows come to Google Cloud Model Garden: GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna roll out in private preview to selected enterprise customers, sold through Google Cloud Marketplace, with charges applicable to existing Google Cloud spend commitments. General availability is expected in the coming weeks, according to the announcement. The signal for your business: buying AI models now looks more like buying software in the cloud you already use than opening a new account with every vendor.
The three models and what each is for
- GPT-6 Astra: demanding work in coding, computer use, professional tasks and scientific research. The top of the family.
- GPT-6.1 Sol: near-Astra intelligence at lower cost, built for agentic coding, computer use and professional work.
- GPT-6 Luna: a fast, low-cost model for everyday work at scale, such as classification, routing and bounded decisions.
- Codex: OpenAI's agentic coding workflows also enter Model Garden, so teams can connect their development tools (CLI and desktop) to inference on Google Cloud.
What changes in practice for a buyer
- One invoice, less buying friction: buying through Marketplace lets OpenAI model spend count against the Google Cloud commitment your company already holds. That removes a classic internal approval barrier: a new vendor with a new contract.
- Your cloud's governance, another vendor's models: Google Cloud says prompts, context and outputs stay inside designated data perimeters and are not used to train foundation models. Zero Data Retention and Confidential Compute options are announced for a later release, so if your use case requires those guarantees, verify them in writing before moving sensitive data.
- Microsoft remains the primary partner: the announcement says so explicitly. Google Cloud is an additional distribution channel, after AWS. Read this as real multicloud, not a change of sides: pick the model per task and the cloud per governance and price.
The honest note
- Private preview for now: only selected customers are in today. If your team is not in the preview, the useful move is to prepare the evaluation (use cases, metrics, budget per million tokens) for when it opens.
- Google Cloud pricing not yet published: OpenAI's public API list prices for Astra, Sol and Luna are a reference, but Google has not published its own prices for these models in Model Garden. Do not budget assuming they match.
- Coming next, as announced: the models will also arrive in Gemini for business and Google Antigravity, Google's coding tool. If your team already lives in that ecosystem, the future integration matters more than today's announcement.
Related links
- What an AI agent is and how a real business uses one
- The 7 types of AI agents for business (and which one yours needs)
- How to implement AI in your business in 30 days (step-by-step plan)
Frequently asked questions
What did OpenAI announce with Google Cloud on October 21, 2026?
That its GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna models, plus Codex workflows, come to Google Cloud Model Garden in private preview for selected enterprise customers, sold through Google Cloud Marketplace, with general availability expected in the coming weeks.
Can I apply my Google Cloud spend commitment to OpenAI models?
According to the announcement, yes: Marketplace purchases can be applied to existing Google Cloud spend commitments. Confirm with your account team how it is counted under your specific contract.
Does this mean Google Cloud replaces Microsoft as OpenAI's partner?
No. The announcement explicitly says Microsoft remains OpenAI's primary cloud partner. Google Cloud is an additional distribution channel, after AWS.
Is my data used to train models?
Google Cloud says prompts, context and outputs stay inside designated data perimeters and are not used to train foundation models. Zero Data Retention and Confidential Compute are announced for later: verify them in writing if your use case requires them.
