OpenAI and Google Cloud announced on 21 October 2026 that OpenAI models are joining Google Cloud's Model Garden. Google says GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are entering private preview for selected enterprise customers. OpenAI plans general availability in the coming weeks, so this is not immediate access for every cloud customer.
The move allows organizations to buy and run OpenAI models within Google's cloud framework, including billing, permissions and data controls. Alongside model access, the companies describe connections to Codex workflows. Commercially, model choice can expand without requiring a separate purchasing arrangement with another vendor for every selection.
Three models with different roles
Google positions GPT-6 Astra for demanding coding, computer use, professional work and science. It describes GPT-6.1 Sol as a lower-cost option approaching Astra's capabilities in agentic work, while GPT-6 Luna targets fast, cost-efficient everyday tasks. These are company descriptions, not results from a new joint benchmark published with the announcement.
These distinctions matter because enterprise applications perform different tasks. A short summary, code repair and complex document review can have different requirements. Switching models within one platform may make matching them easier, but selection still requires testing against the data, language and accuracy needed in each workflow.
Google says Model Garden includes more than 200 models from Google, other suppliers and the open-model community. OpenAI expands that catalog. This does not mean every model behaves identically or that an existing system can move between them without adaptation. Response formats, latency and tool use remain practical considerations.
For developers, a unified route can shorten the distance between experimentation and integration into an existing system. However, inclusion in a catalog does not establish suitability for a particular service. A useful evaluation applies the same task and success criteria to several models instead of relying solely on their commercial descriptions.
Billing through Google Cloud Marketplace
Both companies say purchases will go through Google Cloud Marketplace, with usage appearing on the Google Cloud invoice. Organizations can apply existing cloud spending commitments to these purchases. Google adds that usage counts toward Flexible Savings Plans across model providers, subject to the relevant account plan and terms.
That matters to IT leaders and procurement teams. Budget already allocated to one cloud may become usable for another supplier's model through the same billing channel. It does not guarantee lower overall costs for every organization. Usage volume, tasks and commercial agreements affect the actual bill.
The announcements provide neither a full price list for this route nor a total-cost comparison with buying directly from OpenAI. There is therefore no basis for claiming a uniform discount. The immediate value described is consolidated purchasing and use of existing commitments; financial suitability still depends on terms and deployment.
Consolidated billing also does not mean an unlimited budget. Organizations connecting agents to workplace systems need to understand how often models are called, which tasks trigger more expensive models and how repeated requests accumulate. Model selection becomes an ongoing operational decision, not simply a choice made during installation.
Data controls, and what is still coming
Google says prompts, context and outputs remain within the customer's designated data perimeter and are not used to train foundation models. It describes dedicated Google Cloud infrastructure, isolated access gateways, identity management and unified audit logging. These are supplier commitments and architectural descriptions that must be understood within the applicable service terms.
The announcement says every agent or model interaction operates under an authenticated identity, with VPC Service Controls protection. The aim is to let administrators understand who accesses information and track activity. Having access controls does not replace configuring permissions correctly, particularly deciding which information sources an agent may reach.
Two security capabilities are explicitly forthcoming: Zero Data Retention options and Confidential Compute. The former concerns policies against retaining data; the latter protects information processing within an isolated environment. They should not be described as available to every preview participant already, or as automatically covering every workload.
The distinction between not training on data and not retaining it is substantial. A supplier can promise not to train a model on customer content without that promise defining all retention periods and logs. Organizations with privacy requirements therefore need the specific service policy, rather than treating different questions as one general security claim.
Codex connections now, further integrations later
Google says eligible customers can use APIs and connect local Codex CLI and desktop workflows to its Gemini platform for inference. This connects a working environment to a compute service. It does not announce that every Codex product has moved to Google or that all users automatically receive new access without configuration.
The companies also describe future integration with Google Antigravity and enterprise Gemini tools. OpenAI refers to Gemini Agent, while Google names Gemini for business. Neither announcement provides a detailed timetable for each interface, so the private access opening now must be distinguished from later integrations.
For Israeli organizations already using Google Cloud, the option is particularly relevant to purchasing, development and information governance. However, the announcements do not specify availability in the Israeli cloud region or promise local data processing. Global integration does not establish that the models run in Israel or automatically meet every local requirement of an organization.
OpenAI says Microsoft remains its primary cloud partner. The Google partnership adds a commercial and technical access route without changing that statement. For customers, the tangible change is more choice within a familiar working environment. Its value will depend on availability, costs and controls suited to the actual workload.