OpenAI's GPT-6 models enter private preview on Google Cloud, putting OpenAI on all three major US clouds

The three models, as positioned by the vendors (Unite.AI):

• GPT-6 Astra: the top tier, for coding, computer use, professional work and scientific research.
• GPT-6.1 Sol: close to Astra at lower cost.
• GPT-6 Luna: fast and cheap, for high-volume everyday work.

Access:

• Available now: APIs, plus connecting local Codex CLI and Codex Desktop workflows to Google's platform.
• Coming later: Antigravity (Google's coding tool) and Gemini for Business.
• Access route: an interest form, for selected enterprises only (Digital Applied).

The sources do not cover architecture, training details or benchmarks for these models. What they do document is how the integration works:

• Data path: requests go through isolated gateways. Prompts and outputs stay inside the customer's data perimeter and are not used for training. VPC Service Controls, unified identity and audit logging are available (Unite.AI).
• Not yet available: zero data retention and Confidential Compute (Digital Applied).
• Platform: Model Garden belongs to the Gemini Enterprise Agent Platform, which absorbed Vertex AI in April 2026 and lists 200+ models (Let's Data Science).
• Docs: at publication, the models were not yet listed in Google's public documentation. Teams cannot yet check regions, quotas or API parity.

• Google says it is the only hyperscaler offering all three leading closed models (Unite.AI). This holds because AWS and Azure don't offer Gemini. But on Google, GPT-6 is still in preview, while it is reportedly generally available elsewhere (Digital Applied).
• Model capability tiers come from OpenAI and Google descriptions only. The reviewed sources contain no independent evaluation.
• Spend on OpenAI counts toward Flexible Savings Plans, but without the FSP discount, according to Google's own docs as cited by Digital Applied.
• Cloud growth figures (Google about 82% in Q2, AWS 37%, Microsoft 42%) are from CNBC's reporting (CNBC).

• OpenAI already runs inference on Google infrastructure and has done since 2025 (DCD). Selling models through Google's marketplace is the newer step.
• Model Garden already hosts Anthropic Claude (Let's Data Science).
• Google announced Gemini 4 Argon in September. OpenAI is reportedly seeking a $1.4 trillion valuation and planning an IPO (CNBC).

• Preview is not GA. Commitments that only apply after launch shouldn't be read into the preview. The announcements alone don't support data-residency claims, including for EU compliance (SpazioCrypto).
• Easier buying, not necessarily cheaper. A simpler purchase path doesn't mean a lower price, and no rate card exists yet (SpazioCrypto).
• Microsoft stays OpenAI's primary cloud partner. This adds a channel; it doesn't replace Azure (SpazioCrypto).
• Unconfirmed dates: the date of the Microsoft amendment and the September GA dates each rest on conflicting or single sources.

• Procurement: Google Cloud customers can now buy GPT, Gemini and Claude on one invoice and draw down existing commitments (Unite.AI).
• Multi-model routing: having all three providers in one place reduces the case for building around a single cloud's models.
• Before production: wait for published pricing, region lists, zero data retention and GA terms.
• Developers: Codex users can test the Google-hosted backends now. Antigravity integration is still pending.
• Market context: Ramp data shows most business AI spending already goes to OpenAI and Anthropic (CNBC). The move gives Google Cloud access to that spend, even when it isn't going to Gemini.