On October 21, OpenAI and Google Cloud announced that OpenAI’s models and its Codex coding agent will be available to Google Cloud customers. That makes OpenAI a multi-cloud vendor in practice: it is already on Azure and AWS, and Google was the last of the three to sign.
What was announced
- Where: OpenAI’s models go into Google Cloud’s Model Garden, alongside more than 200 other models. Codex is included.
- When: “In the coming weeks.” One report says the models start in private preview for select enterprise customers, with broader availability, including Gemini for business and Google’s Antigravity coding tool, expected later.
- Which models: That report names three tiers: GPT-6 Astra (flagship), GPT-6.1 Sol (cheaper, agentic coding) and GPT-6 Luna (fast and low cost). Other coverage only names GPT-6 Astra, so treat the full lineup as reported, not confirmed.
- How you buy it: Through Google Cloud Marketplace. Spend appears on the normal cloud invoice and counts toward existing cloud commitments.
- Not disclosed: Pricing and any revenue share between the companies.
Why it matters
Until this year, Azure was the main official route to OpenAI models at scale. In April, Microsoft and OpenAI restructured their deal so Microsoft’s license is no longer exclusive; Microsoft keeps a revenue share and IP access through 2032. Amazon then signed a $50 billion partnership with OpenAI. With Google now in, all three hyperscalers sell OpenAI models, and Anthropic’s Claude is also available on all three.
The practical change is procurement. If your cloud commitment is with Google, you can now spend it on OpenAI models without a separate contract. The same logic applies on Azure and AWS.
Where to be skeptical
- Preview is not production. Access is limited at first, and “coming weeks” has no date.
- No price, no comparison. Per-token pricing on Google may differ from OpenAI direct, Azure or AWS. Wait for published rates.
- Feature parity is unproven. Hosted versions of a model do not always get every feature or new release at the same time.
- Competitive motives. Each cloud wants model choice as a reason to keep your commitment. That is good for buyers, but the lock-in just moves to the platform layer.
What to check before you move anything
- Confirm whether your existing commitment can be drawn down through the marketplace, and at what rate.
- Ask where inference runs and where prompts and logs are stored, especially if you have Canadian data-residency or privacy requirements.
- Compare latency, quotas and rate limits across clouds for your real workloads.
- Keep an abstraction layer between your apps and any one model API so you can switch providers.
- Review data-handling terms: Google says prompts stay out of model training and highlights data perimeters and audit logging.
