Vertex AI is the Google Cloud environment teams use to build, train, deploy, and manage machine-learning models and AI applications. If your team is weighing AI adoption, its familiar name points to managed tools for enterprise teams, not to one model. Google Cloud now presents its broader AI environment as the Gemini platform, a name that can make older instructions harder to follow.
Here, you’ll get a practical look at its tools, model choices, workflows, uses, and cost structure. You’ll also see where current product names differ from names you may find in older guides. That matters when you compare services or follow setup steps. By the end, you should be able to tell whether its managed tools fit your team’s AI work.
Key Takeaways
- Vertex AI is a familiar name for Google Cloud’s managed machine-learning environment.
- Gemini models are not the same thing as the broader Gemini platform.
- Model Garden helps you compare models and tools.
- Training options range from managed jobs to custom code.
- Costs depend on model use, compute, and other Cloud services.
What Is Vertex AI?
Vertex AI brought Google Cloud’s machine-learning development and deployment capabilities into a managed environment when it launched in 2021. It gives you one place to build and run machine-learning and generative AI applications. A managed platform groups tools and computing resources, but your team still chooses its models and approach.
Vertex AI launched in May 2021 to support machine-learning model development and deployment. The platform brings services for model work together, so teams can build without setting up every piece of infrastructure on their own. Managed does not mean automatic: your team still selects data, training methods, and deployment needs.
Choose the model, training approach, and deployment method that fit the task; a managed platform does not make those choices for you.
Google Cloud now uses Gemini platform as its broad name for a setting to build, scale, govern, and improve enterprise AI agents. It also supports training, testing, and tuning machine-learning models. That wider product name describes an environment, not a single model.
Gemini models are one family of models available in Google Cloud’s AI offerings. Gemini platform is the broader environment for working with models, agents, and machine-learning applications. Mixing up the two can lead you to follow the wrong setup instructions or expect a model name to describe an entire service.
The name Vertex AI remains common in product searches and older guides. Product labels can change while related tools and instructions are updated at different times. Treat Vertex AI as the familiar search term for the platform’s ML capabilities, and check current Google Cloud product labels when following instructions.
How Does a Machine-Learning Workflow Run on the Platform?
A project usually starts by preparing and checking the data that will teach or inform a model. You can explore it in notebooks, connect it to training tools, and then test whether a model works for the task. MLOps means the practices and tools used to automate, track, deploy, and monitor machine-learning systems.
Google Cloud’s Gemini platform includes Colab Enterprise and Workbench notebooks, with native BigQuery integration. BigQuery and Cloud Storage can connect data to machine-learning work. These tools let data scientists examine records, shape inputs, and try code before committing to a longer training job.
Prepare and connect training data
Start by checking whether the data reflects the question your application needs to answer. You might remove errors, select useful fields, or split examples for training and testing. A model trained on mismatched examples may learn patterns that do not help with real predictions.
Notebook-based development gives you a place to inspect data and write code. BigQuery can hold and query analytic data, while Cloud Storage can store files used by a project. Their connections to machine-learning tools can cut down on manual transfers between separate systems.
Train with managed or custom jobs
For some work, you can use a managed training option rather than running each task on your own machines. You can also write custom code with open-source frameworks such as TensorFlow and PyTorch. That route gives your team more control over the model training process and its framework.
A custom job still needs working code, a suitable machine, and training data. Google’s hands-on lab shows how to package TensorFlow training code, submit a custom job, save a model to Cloud Storage, and serve it. It is a custom training and serving lab that makes the steps concrete.
Evaluate results before deployment
Training is not proof that a model is ready. Use test data and task-specific checks to judge whether its predictions are useful, consistent, and safe for the intended application. For a classification task, for example, you may need to inspect which errors matter most.
A simple project can follow three steps:
- Prepare the data and check that it matches the task.
- Train a model, then evaluate its results against clear criteria.
- Register the model or refine it if results fall short.
That sequence gives a team a record of what it tested and what changed. Keep training data and evaluation criteria aligned with the prediction the application actually needs.
Which Models and Training Options Can You Use?
Model Garden is the place to compare Google, third-party, and open models within Google Cloud’s AI environment. The Vertex AI model list spans different kinds of models and tools, so a catalog entry is a starting point, not a recommendation. Check each model’s launch status and supported inputs before you build around it.
Model Garden offers more than 200 Google and third-party AI models and tools. Its catalog groups offerings as foundation models, fine-tunable models, and task-specific solutions. The Vertex AI models in those groups can differ in access, input types, and readiness for use.
Browse foundation, fine-tunable, and task-specific models
Foundation models can handle a range of tasks, while fine-tunable models support customization for a more focused job. Task-specific solutions target a narrower need. The Model Garden catalog helps you sort through those options before you plan tuning or serving.
| Model Garden group | What it contains | Typical next step |
|---|---|---|
| Foundation models | Models for broad tasks | Test prompts and outputs |
| Fine-tunable models | Models that support customization | Check tuning support |
| Task-specific solutions | Tools for focused tasks | Match the tool to the task |
Compare Google, partner, and open models
Examples include Gemini 4 Argon, Gemma, and Anthropic’s Claude family. Gemini 3.1 Pro entered Model Garden preview on February 19, 2026. It has a 1-million-token context window and accepts text, audio, images, video, PDFs, and code repositories.
Claude Sonnet 4.6 became available on February 17, 2026, and Claude Opus 4.7 followed on April 14, 2026. Gemma 4 26B A4B IT entered experimental launch as a managed API on April 3, 2026. It accepts text and image input and generates text output.
Availability labels matter: preview, experimental launch, and general availability describe different stages. Google Cloud scans its supplied model-serving and tuning containers for vulnerabilities. It also scans featured partners’ third-party checkpoints for authenticity.
Choose a managed API or custom training
You can select an existing model, customize a supported model, or use custom model training. Custom training gives you control over training code and framework, but your team takes on more setup. An existing model can be a better start when its task and input types already fit.
After choosing, the common path connects a model with tuning, evaluation, and serving tools. If you are comparing a model available from another provider, a Mistral Large 4 API setup guide can help you assess that separate route; it does not mean the model is in Model Garden. Before adopting a model, verify its launch status and confirm that its input and output types fit your task.
How Do Teams Deploy and Monitor Models?
A model is ready for application use only after it has been registered and connected to a prediction method. You can request online predictions through a deployed endpoint or run batch prediction on groups of records. The right method depends on when the application needs its results.
Model Registry helps manage model versions, while Pipelines can make workflows repeatable. Keep a clear record of which model version serves each application. That record helps your team track changes when a model or workflow is updated.
Serve predictions online or in batches
An endpoint handles online prediction requests as an application needs them. Batch prediction processes a larger group of records together, which can suit tasks that do not need an instant answer. Your application’s response needs help determine which approach fits.
A registered model can support either kind of prediction. Serving online also means keeping the endpoint available for requests, while batch work runs over a set of records. Both methods connect a model to an application, but their timing differs.
Track model versions, workflows, and drift
Pipelines can organize repeatable work, such as a sequence of model tasks. Feature Store supports sharing and reusing machine-learning features across projects. Shared features can help teams use the same prepared inputs rather than rebuild them separately.
Monitoring can detect input skew and drift. Input skew means the data reaching a model differs from the data used during training. Drift means patterns in the data or model results change over time. Those signals can alert a team that a model may need review.
Monitoring only helps when someone acts on its signals. Set a monitoring owner and a response process before relying on a deployed model in production.
What Can AI Agents and Search Applications Do?
A generative AI agent can use a model together with enterprise information and tools to complete a workflow, rather than only return a response. An agent may access data, follow instructions, and call tools to carry out a task. Search applications serve a related but distinct role by finding information and forming answers.
The open-source Agent Development Kit supports programmatic agent development, including agents that use non-Google models. Agent Search is the name Google Cloud uses in its documentation for the service previously called Vertex AI Search. The name differs by surface: documentation says Agent Search, while the console and APIs retain older labels.
Build agents with frameworks and enterprise data
Agents combine a model with instructions, tools, and data access. For example, an agent can search business information and then use a tool to complete a related step. A Google Codelab shows agents working with Google Workspace data, custom tools, and the Google Workspace agent integration lab.
Vertex AI Extensions was deprecated on May 26, 2026, with shutdown scheduled after November 26, 2026. Google Cloud recommends migration to Agent Platform. Teams using the older extension should plan around that change rather than treating it as a current long-term option.
Use Agent Search for retrieval and answers
Agent Search retrieves information that can support answers in a search application. Its documentation name changed from Vertex AI Search in 2026, while the console kept the Vertex AI Search and AI Applications name. APIs retained Discovery Engine endpoints, so code and interface labels may not match.
Track the 2026 search and RAG changes
Google Cloud made several changes to Agent Search and retrieval-augmented generation, or RAG, which connects model answers to retrieved information. The dates and release stages help show what is ready, in preview, or being retired.
- April 3: Vertex AI RAG Engine serverless mode entered public preview, with managed storage that does not require user-provisioned database scaling.
- April 6: RAG Engine added schema-based metadata search. April 17: RAG Cross Corpus Retrieval entered public preview for retrieval or answer generation across corpora through AsyncRetrieveContexts and AskContexts.
- April 22: Agent Search’s hosted Model Context Protocol server became generally available at discoveryengine.googleapis.com/mcp. Document-level relevance and semantic-relevance threshold filters also became generally available.
- May 13: Prefix and partial matching for Agent Search filters entered public preview, along with the EXISTS filter for custom and media search. May 15: Agent Search for healthcare was deprecated; custom Agent Search apps or Agent Retrieval were suggested for users needing more retrieval control.
- May 27: Layout parser table and image annotations became generally available. They can describe table or image information used as sources in generated answers.
- August 13: Gemini 3.5 Flash was added for answer generation. September 14: Customer-managed encryption-key rotation became generally available for Agent Search apps and connectors, with automatic re-encryption when a new key version becomes primary in Cloud KMS.
- October 16: Agent Search changed its stable answer-generation model from gemini-2.5-flash/answer_gen/v1 to gemini-3.5-flash/answer_gen/v1.
These changes affect different parts of search and RAG, from answer generation to data retrieval and encryption. Check the current service name, API, and feature status before updating an existing search or agent integration.
How Does Vertex AI Pricing Work?
Vertex AI costs are built from the model, compute, and services a project actually uses. Charges can also include storage and other Google Cloud resources. Actual cost depends on the tools and workload you run, not just the model name.
Open-source Model Garden tuning is billed for compute resources at custom-training rates. Deploying a model is billed for the compute resources used to run its endpoint. Custom-training charges depend on machine type, region, and accelerators.
Understand what each usage charge covers
Listed text, chat, and code generation starts at $0.0001 per 1,000 characters. Pipelines pricing starts at $0.03 per run. These starting prices describe specific charges, not the full cost of a project.
| Charge | How it is billed | Published starting point or estimate basis |
|---|---|---|
| Text, chat, and code generation | Per 1,000 characters | $0.0001 per 1,000 characters |
| Pipelines | Per run | $0.03 per run |
| Custom training | Compute resources used | Machine type, region, and accelerators |
A tune or deployment can add compute charges beyond model use. Your estimate should account for how long the endpoint runs and how much training compute each job needs. Other Cloud services used by the application can add their own charges.
Check free credits and estimate a project
New Google Cloud customers get $300 in free credits. Those credits can help pay for trials, but they do not make all Vertex AI use free. Usage beyond the credit is billed under the rates that apply to each service.
Start an estimate by listing the model calls, training jobs, endpoint runtime, pipeline runs, and related resources your project expects. A short test can reveal which usage items matter most before you set a larger budget. Estimate the expected training runs, endpoint runtime, and model usage before approving a production budget.
What Should You Check Before Choosing Vertex AI?
Vertex AI is most compelling when its managed tools match the team’s existing data and operational needs. It can suit teams that want Google Cloud integration and managed machine-learning operations. Fit comes first: a strong tool can still be a poor match for your stack.
Gemini platform supports open-source frameworks for training and deployment, and it lets customers choose AI infrastructure for training and prediction. Colab Enterprise and Workbench offer notebook options integrated with BigQuery. Your team can make framework and infrastructure choices within the platform.
Match platform capabilities to your team
Use this checklist to compare the platform with the work your team already does:
- Where does the data already live?
- Does the team need managed training or custom training?
- Which machine-learning frameworks does the team use?
- What infrastructure does the application need for prediction?
- Would Colab Enterprise or Workbench fit the team’s notebook workflow?
BigQuery integration can help when useful data already sits in BigQuery. Notebook choices also matter: data scientists may prefer a notebook-based process, while another team may put more weight on custom code and deployment options.
Confirm infrastructure and migration needs
Consider how a new platform would connect with current systems, data access rules, and the people who will operate the service. Google Cloud can be a strong choice for data and AI workloads, but fit depends on your existing systems, application needs, and team skills. A cloud platform comparison can help frame those trade-offs.
Vertex AI is a better fit when your team wants managed ML operations and Google Cloud integration. It is a weaker fit when those strengths do not match your stack or workflow. Validate one representative workflow and its expected resource use before expanding adoption.
What Are Common Vertex AI Use Cases?
The most useful applications connect a model to a specific task, information source, and workflow. Teams can build customer and employee support, knowledge access, creative and code tools, data access, and security response. Start with a real task, not an agent looking for a job.
Google Cloud’s April 2026 customer-use-case article said nearly 80% of its customers were using its AI products, and nearly 90% of Fortune 100 companies were using Gemini Enterprise. Those adoption figures describe use, not proof that a tool will suit your team or produce a particular result.
Choose a use case with a clear owner and a measurable business task before expanding to a broader agent program.
Support customer and employee work
Customer agents can handle frequent questions and support contact-center work. A voice agent can help manage spoken requests, while an FAQ agent can respond to common questions. Employee agents can help with tasks such as invoice processing or finding internal documents.
Knowledge agents can answer questions based on company information. An internal-document assistant, for example, can help staff locate details in business files. These applications depend on access to useful information and a clear task for the agent.
Automate creative, data, and security workflows
Creative agents can support generative-media production, while code agents can help with software tasks. Data agents can let users ask questions in natural language about older systems. This can make legacy data easier to reach without asking every user to know the system’s query language.
Some applications use multimodal analysis, which means a model can work with more than one kind of input. A team might analyze physical environments using images and other information. Security agents can help automate parts of a cybersecurity response, such as reviewing alerts and starting a defined workflow.
Google Cloud’s April 2026 use-case list includes customer, employee, creative, code, data, and security agents. A separate 2026 guide groups agent use cases as customer, employee, knowledge, and voice agents. Its examples include FAQ handling, invoice processing, internal-document assistants, and contact-center agents.
Use the adoption figures as context, not a substitute for testing your own task. Choose a use case with a clear owner and a measurable business task before expanding to a broader agent program.
Conclusion
Vertex AI brings model development and operational tools together for teams building AI applications on Google Cloud. It remains a familiar name for a broad managed AI environment, while Google Cloud uses Gemini platform for its wider platform branding.
Product names can differ across documentation, consoles, and APIs, so check the label for the specific tool your team plans to use. The right choice depends on your data, model needs, and the infrastructure your application requires.
If you want more practical coverage of AI and technology, you can choose topics through the Occcy newsletter. Pick one concrete use case, map its data and deployment needs, and estimate the resources required for a small proof of concept.
FAQ
What does Google use AI for?
Google Cloud customers use AI for customer, employee, creative, code, data, and security applications. Examples include FAQ handling, invoice processing, internal-document assistants, contact-center agents, media production, and cybersecurity response. These are customer use cases supported by Google Cloud products. They do not, by themselves, prove how Google uses AI inside its own operations or predict results for your team.
Is Vertex AI gone?
No, Vertex AI remains a familiar name for Google Cloud’s machine-learning capabilities. Google Cloud uses Gemini platform for its broader AI platform, while documentation renamed Vertex AI Search to Agent Search in 2026. The console retained the Vertex AI Search and AI Applications name, and APIs kept Discovery Engine endpoints. Those naming differences can appear in the same setup process.
Is Vertex AI the same as Gemini?
No, Vertex AI and Gemini are not the same thing. Gemini is a family of models, while Gemini platform is Google Cloud’s broader environment for working with models, agents, training, and deployment. A Gemini model may be used within that environment, but the platform includes more than that model family. Check whether a product label names a model or a service.
Can I use Vertex AI for free?
New Google Cloud customers get $300 in free credits that can help cover Vertex AI trials. The credit is not a promise that every service or amount of usage is free. Once the credit is used, charges apply to the model, compute, and other Cloud resources your project uses. Track your trial use so you know when charges may begin.
How much does Vertex AI cost?
Pipelines pricing starts at $0.03 per run, while other charges depend on model use, compute, region, and configuration. Listed text, chat, and code generation starts at $0.0001 per 1,000 characters. Training and endpoint use can add compute charges, and other Cloud resources can add more. Estimate each part of your project instead of treating one rate as its total cost.
Does Google use Vertex AI?
Google Cloud provides Vertex AI capabilities to customers who build and operate machine-learning applications. That fact does not establish which tools Google uses internally for its own work. Customer examples show what teams can build with the platform, not what Google itself has deployed. Keep those claims separate when you assess the platform or discuss its use.
Is Vertex AI worth it?
Vertex AI is a better fit for teams seeking managed machine-learning operations and Google Cloud integration. Teams with different infrastructure, data systems, or workflows should compare alternatives and test a representative task. The value depends on the tools your project needs and the resources it uses. A small proof of concept can show whether the platform fits before you expand adoption.
