OpenAI on Google Cloud: Preview, Billing and Codex
OpenAI models are coming to Google Cloud Model Garden. See who has preview access, how Marketplace billing works, and what is still coming to Codex and Antigravity.
OpenAI and Google Cloud announced a new distribution partnership on October 21, 2026. Enterprises will be able to use OpenAI frontier models through Google Cloud's Model Garden and purchase access through Google Cloud Marketplace. The headline is model choice inside an existing cloud governance and billing environment—not that every Google Cloud account can immediately switch on GPT-6 Astra. Google says the initial rollout is a private preview for selected enterprise customers; OpenAI says general availability is planned in the coming weeks.
What is available now, and what is coming?
Google identifies three incoming models: GPT-6 Astra for difficult coding, computer-use and professional tasks; GPT-6.1 Sol for strong agentic work at lower model cost; and GPT-6 Luna for faster, cost-sensitive high-volume tasks. These are workload descriptions, not a guarantee that one model wins on your data. Google says selected preview customers can build through its APIs now and connect local Codex CLI and Desktop workflows to the Gemini platform for inference. Broader model access, OpenAI models inside Google Antigravity, and Gemini for business integrations are described as coming soon.
Private preview is not general availability. OpenAI's announcement targets GA in the coming weeks, while Google Cloud describes current access as selected-enterprise preview. Confirm your organization and region before scheduling a production migration.
Why this matters for enterprise teams
The practical benefit is procurement and operations consolidation. Google says OpenAI usage bought through its Marketplace can appear on a Google Cloud invoice, use eligible existing cloud commitments, and count toward Flexible Savings Plans. For a team already operating agents in Google Cloud, that could reduce separate purchasing and governance work. It does not prove lower token prices or automatically make every workload cheaper: compare the actual commercial terms, model quality, latency, throughput and required safeguards before committing.
A first-pass model selection guide
| Workload need | Model to evaluate first |
|---|---|
| High-stakes reasoning, complex coding or research | GPT-6 Astra; test whether additional capability earns its cost |
| Multi-step coding agents and professional workflows with cost constraints | GPT-6.1 Sol; measure success and escalation rates |
| Frequent, simpler operations where speed and unit cost dominate | GPT-6 Luna; check task-level quality before scaling |
Governance claims still need validation
Google says prompts, context and outputs remain within the designated data perimeter and are not used to train foundation models. It also describes isolated infrastructure, authenticated identities and audit controls. But two details are explicitly future features: Zero Data Retention options and Confidential Compute. Do not mark those as already deployed, and do not treat a provider's general security description as proof that your specific workflow satisfies residency or regulatory obligations. Test actual data paths, logging, identities and contracts in the preview.
A sensible evaluation plan
- Request access and verify the exact models, region, endpoints, billing terms and quota available to your organization.
- Choose three representative tasks—one difficult, one agentic, one high-volume—and set measurable success, latency and cost thresholds.
- Test the same tasks on your current model stack and on eligible preview models, including failures and retries.
- Review data perimeter, audit logs, retention settings, user permissions and procurement with security and finance teams.
- Keep a fallback model and do not announce a migration date until the intended access tier is confirmed.
For developers, the timing distinction is particularly important. Google announced Claude Opus 5.5 and Sonnet 5.5 as available in Antigravity on October 20, with administrator enablement. Its October 21 announcement describes OpenAI models in Antigravity as upcoming. The two announcements should not be combined into an unsupported claim that GPT-6 models already appear in every Antigravity model picker.
Questions people ask
Can I use GPT-6 Astra on Google Cloud right now?
Only selected enterprise customers are being admitted to Google Cloud's private preview as of October 21, 2026. Google lists GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna for that rollout and offers an interest form. OpenAI says general availability is planned in the coming weeks, but neither announcement gives a guaranteed activation date for every account or region. Verify eligibility before planning production use.
Can an existing Google Cloud spending commitment cover OpenAI models?
Yes, the companies say eligible Google Cloud commitments can be applied to OpenAI model purchases made through Google Cloud Marketplace. Google also says the usage appears on a Google Cloud invoice and counts toward Flexible Savings Plans. That is a procurement and billing arrangement, not a published promise of lower per-token prices. Confirm contract eligibility, regional availability, negotiated rates and any preview restrictions with your account team.
Are OpenAI models already available inside Google Antigravity and Codex?
Not as a broadly available Antigravity model-picker option in the October 21 announcements. Google says OpenAI models in Antigravity and Gemini for business are coming soon. It separately says selected preview customers can connect local Codex CLI and Desktop workflows to its Gemini platform for inference through available APIs. Those are different access paths, so check which one your organization actually has.
Will Google Cloud retain or train on prompts sent to OpenAI models?
Google states that customer prompts, context and outputs for these workloads stay within the designated data perimeter and are not used to train foundation models. However, the same announcement says Zero Data Retention options and Confidential Compute are coming soon, rather than universally active today. Review the applicable service terms, logging, regional settings and your organization's security requirements before sending sensitive production data.
OpenAI's arrival in Google Cloud is meaningful for enterprise model choice, consolidated purchasing and agent development. The immediate action is a controlled preview evaluation—not a blanket GA rollout, guaranteed savings, or an assumption that future retention and Antigravity features are already live.
Explore related Make Better resources
- GPT-6.1 Sol changes the agent cost equation— Model economics and workload routing context
- GPT-6 Sol vs Luna: pricing and workflow fit— Complementary model tier selection context
Sources & useful resources
- OpenAI: OpenAI models are coming to Google Cloud— Primary OpenAI announcement, October 21, 2026; planned GA, Marketplace, Codex and Antigravity timing
- Google Cloud: Expanding model choice with OpenAI— Primary Google Cloud announcement, October 21, 2026; three models, private preview, API/Codex details, governance and billing
- Google Cloud: Model Garden— Official platform overview; existing third-party model selection and deployment
- Google Cloud: Claude 5.5 in Antigravity— October 20 primary source clarifies that Claude availability in Antigravity is current while OpenAI Antigravity integration is future