Google Cloud is adding GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna to its model garden, with a private preview rolling out now for select enterprise customers. Eligible customers can use the models through APIs starting today and route inference from their local Codex CLI and Desktop workflows to Google's Gemini platform. Access initially covers selected enterprises, while availability in Gemini for business and Google Antigravity remains forthcoming.
Customers purchase OpenAI model usage through Google Cloud Marketplace, putting those charges on their Google Cloud bill. That spending also reduces existing cloud commitments and qualifies toward Flexible Savings Plans (FSP), including spending across model providers. Google says customers can switch models according to cost, response time and performance without signing separate contracts with vendors or becoming tied to one architecture.
Read more: Google Begins Private Testing for Gemini Mac App with Desktop Intelligence
Google positions GPT-6.1 Sol as a less expensive option with intelligence approaching Astra for professional work, agentic coding and computer use. GPT-6 Astra targets more demanding tasks, including scientific research, while GPT-6 Luna is intended to handle everyday work at scale with an emphasis on speed and cost. The additions join a model garden offering more than 200 models from Google, outside providers and open model projects.
Google says customer inputs, supplied context and generated responses stay inside the customer's designated data perimeter and are excluded from foundation model training. The OpenAI models operate on dedicated Google Cloud infrastructure, with requests passing through isolated gateways that Google says keep transit inside its environment. Zero Data Retention (ZDR) options and Confidential Compute are still forthcoming, putting those protections on a different timetable from the private preview.
For administrators, Google says each interaction involving an agent or model has an authenticated identity. It also provides centralized audit logs and protection through VPC Service Controls across enterprise operations. Google says the access boundaries prevent outside model providers from gaining unauthorized entry to customers' cloud environments.
The expansion follows Google's release yesterday of TorchTPU as open source on GitHub, extending its infrastructure options alongside today's model additions. That earlier release made TPUs a native PyTorch option alongside CPUs and GPUs, allowing customers to train, serve and deploy models using existing tools.
