What does it mean that OpenAI models are now on Google Cloud, and should a small business use them?

Announced October 22, Google Cloud added OpenAI’s GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna to Model Garden in a private preview for select enterprise customers — one Google Cloud bill, prompts kept inside your own data perimeter, and no training on your data. It is not open to everyone yet, so a small business’s job today is to learn its cost per job before committing.

Short answer: as of October 22, a company that already runs on Google Cloud can rent OpenAI’s newest models — GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna — through Google’s Model Garden, bought on the same Google Cloud bill. It is a private preview for select enterprise customers, not a public launch, so most small businesses cannot sign up today — but it shows exactly where buying AI is heading. Why it matters: until now, using OpenAI’s best models usually meant a separate account, a separate bill, and your prompts travelling outside the cloud your company already trusts.

Inside Model Garden, purchases count toward your existing Google Cloud commitments, prompts and outputs stay inside your own data perimeter, and they are not used to train foundation models. Zero Data Retention and Confidential Compute are planned for later — planned, not live yet.1. Match the model to the job.

Astra is aimed at the most demanding work, Sol at balanced everyday jobs, and Luna at fast, lower-cost tasks — the win is not paying top-tier prices for simple work.2. Price the job, not the model. Run one real task — say, drafting 100 support replies — on two models and compare the cost per finished, usable answer, including retries.3.

Check the perimeter in writing. Where are prompts stored, how long are logs kept, and what changes when Zero Data Retention arrives?4. Let developers test cheaply.

Codex CLI and Desktop workflows can connect, so a coding team can trial the models without rebuilding its tooling.5. Keep your exit open. Preview pricing and terms change; keep prompts and data portable so you can switch at general availability.Pro tip: The common mistake is committing a large cloud spend to a preview.

Measure cost per completed job on one small, real workload first — that number, not the model’s name, is what you are actually buying.Ad’s take: One bill, several engines under it — that is how cloud has always worked, and AI is just catching up. If you are small, today’s move is not buying; it is knowing your numbers so you are ready when the preview opens.