Gemini Pro reaches Vertex AI in preview: model, platform and product are not the same
Google opened Gemini Pro and Pro Vision in Vertex AI on December 13, 2023, still in preview. Separating model, platform and product reveals what was actually available and what remained a promise.
On December 13, 2023, Google made Gemini Pro and Gemini Pro Vision available to Vertex AI customers. The decisive qualification was the release stage: the official Vertex AI release notes recorded both models as being in preview, not generally available. For an organization, that word separates an opportunity to experiment from a decision to entrust a stable production process to the technology.
The announcement did not mean the entire Gemini family had reached the cloud, either. Google had introduced three sizes—Ultra, Pro and Nano—but Vertex AI was opening Pro for text tasks and Pro Vision for multimodal inputs. Ultra remained limited to early experiments with a selected group. The useful way to read the launch is to separate three layers that marketing often blends together: the model that produces a response, the platform used to operate it and the specific product a user can access.
First layer: which model was available
Gemini was not a single model. In its introduction to the Gemini 1.0 family, Google described Ultra as the version for highly complex tasks, Pro as the option for scaling across a broad range of tasks and Nano as the efficient on-device version. The Vertex AI access announced a week later was for Pro, not Ultra. Results promoted for Ultra therefore could not automatically be assigned to the model a developer could call through the API.
That distinction matters whenever a comparison table is presented. A family name can cover variants with different capabilities, costs and intended destinations. Before accepting a sentence such as “the model beats another model,” ask which variant was evaluated, under what configuration, and whether that is the same variant being offered. Here, the launch page tied several headline results to Ultra, while the December 13 enterprise release centered on Pro.
Model capability must also be separated from the modality exposed by a service. The Google Cloud announcement for Vertex AI described applications working across text, code, images and video. The release notes, which were more precise about what could be invoked, listed Gemini Pro and Gemini Pro Vision separately. Operational documentation takes priority over a broad adjective: an application design needs the model identifier, supported inputs and release stage for each function.
Second layer: what Vertex AI added
A model generates outputs; a platform organizes what is needed to use it within a system. Vertex AI provided API access, identity and data controls, evaluation tools, safety filters and ways to connect a model to outside information or actions. Google said customers would remain in control of their data and that their data would not be used to train its models. Those properties belonged to the Google Cloud wrapper, not to additional intelligence inside Gemini Pro.
The distinction prevents two opposite mistakes. One is assuming that a good platform automatically corrects a false answer. The other is reducing the entire assessment to a model benchmark and overlooking controls that materially affect risk: who may submit data, what is logged, which sources are retrieved, which content is blocked and how an application is observed after deployment.
The launch mentioned customization techniques, data retrieval, extensions and an automated model-comparison tool called Auto SxS. The described possibilities were not all at the same maturity level. Some were existing Vertex AI capabilities; others were future integrations. The announcement, for example, said Gemini Pro was “coming soon” to summarization and answer-generation features in Vertex AI Search, while its role as a foundation model for voice and chat agents would be in preview. A roadmap is not a list of functions ready on launch day.
Third layer: preview is not a production guarantee
The word preview does not make a tool useless. It means the tool should be tested within boundaries, with non-critical data and verifiable outputs, before stability is assumed. A minimum assessment records the version and region, builds a set of representative examples, measures failures that matter to the use case and defines what happens when the model sounds confident without support.
Google acknowledged that boundary. Its Gemini introduction said work was continuing on factuality, grounding, attribution and corroboration. The Gemini 1.0 technical report added that language models still generate hallucinations and struggle with causal understanding, logical deduction and counterfactual reasoning. Its model card warned that Gemini should not be used in downstream applications without first analyzing potential harm in the particular application.
Safety filters and moderation do not replace that assessment. They may reduce some classes of content, but they do not prove that an answer is true, that financial guidance is appropriate or that a summary preserved an important exception. A useful test resembles the actual work: documents of the same type, ambiguous requests, incomplete data and examples where the correct response is to acknowledge that information is missing.
Two entry points with different contracts
On the same December 13 date, developers could access Gemini Pro through Google AI Studio and the Gemini API, or use Vertex AI. Google described AI Studio as a free web tool for prototyping and obtaining an API key. Vertex AI was the managed route for organizations needing data control and enterprise security, privacy, governance and compliance features. Sharing a model family did not make the two environments interchangeable.
The legal protection in the announcement also required careful reading. Google Cloud described a two-pronged indemnification policy, but said coverage for Gemini API outputs was planned for when that API reached general availability. On December 13, that future coverage could not be represented as a guarantee already applying to every Gemini use.
How to read the next launch
Four questions organize the investigation of a similar announcement. First, which model identifier and variant can I use today? Second, does the operational documentation call it preview or general availability? Third, which capabilities belong to the model and which come from the serving platform? Fourth, which limitations does the technical report acknowledge, and how will I reproduce them with my own data?
Then build a small test. Choose a task with a checkable result, retain inputs and outputs, compare at least one alternative and have an accountable person review failures rather than only successes. If the system needs private information, verify that every answer points to the retrieved passage. If it can execute actions, restrict permissions and require confirmation for irreversible ones. The label “agent” describes an architecture; it does not confer judgment or authority.
The transferable skill is straightforward: separate model, platform and product before evaluating a promise. Gemini Pro could be accessible on December 13 and still be in preview; Vertex AI could add enterprise controls without turning every output into a fact; and an attractive roadmap could remain unavailable. Reading those three layers turns a launch into a decision that can be tested.
This article was produced with artificial intelligence under human editorial oversight.
