Google launches Gemini 3 Pro for building AI apps

Google introduces Gemini 3 Pro, an AI model designed to write code, operate tools, and turn instructions into working applications. It is already available in preview for developers through the Gemini API, Google AI Studio, and Vertex AI.

The company describes it as its most capable model to date and says it outperforms Gemini 2.5 Pro in programming, complex tasks, and agent workflows. These capabilities allow AI to do more than respond: it can plan steps, use tools, and carry out actions with a degree of autonomy.

From an idea to an application

Gemini 3 Pro embraces what is known as “vibe coding”: describing what you want to build in natural language and letting the model generate much of the code. For example, you can ask it for a simple game, an interactive page based on voice notes, or an application built from a sketch.

Google says the model scored 1487 Elo points in WebDev Arena, a ranking that compares the quality of web applications generated by different models. The figure reflects a specific evaluation, not a guarantee that any application created with a single instruction will be production-ready.

The feature is available in Google AI Studio’s Build mode. This environment automatically connects the models and APIs needed to add AI features to an application and lets you iterate on the result with additional instructions.

An agent that works in the editor, terminal, and browser

Alongside Gemini 3, Google introduces Antigravity, an agent-based development platform in public preview and available at no cost. It is available for macOS, Windows, and Linux.

The idea is for you to act as the architect while agents handle specific tasks, such as:

• Creating a new feature.
• Modifying an application’s interface.
• Finding bugs and suggesting fixes.
• Researching information and preparing reports.
• Running processes from the terminal or browser.

Agents can work across different environments and document their progress through detailed results. You can also use Gemini 3 Pro from Gemini CLI, Android Studio, Cursor, GitHub, JetBrains, Manus, and Cline, among other tools.

Google says the model scored 54.2% on Terminal-Bench 2.0, a test that measures the ability to use tools from a terminal to operate a computer. This suggests Gemini 3 Pro is designed for more than suggesting lines of code: it can also participate in multi-step development processes.

New tools for the API

The Gemini API adds a bash tool that allows the model to propose shell commands. It can be used to navigate local files, automate development tasks, or run system operations, although applications will need to carefully control which commands are authorized.

A version hosted on Google’s servers is also arriving for generating code in several languages and creating prototypes more securely. This feature is currently available to partners with early access and will become generally available later.

In addition, the Grounding with Google Search and URL context tools can be combined with structured outputs. In practice, an agent can access a page or search for information, extract the relevant data, and return it in a specific format for another system to use.

It understands text, images, and video

Gemini 3 Pro includes a 1 million token context window, meaning it can process large amounts of information in a single interaction. Google also says it has improved at understanding images, documents, and videos.

This enables use cases such as:

• Analyzing complex documents beyond simply reading their text through OCR.
• Understanding annotations and elements on a screen to operate a computer.
• Following fast actions in a video and locating specific details afterward.
• Interpreting positions, trajectories, and movements in robotics, extended reality, or autonomous vehicle scenarios.

The API adds more precise controls for adjusting image resolution and balancing quality, speed, and cost. It also introduces a new reasoning level and stricter validations to preserve the model’s internal process during multi-turn conversations.

Pricing and availability

Gemini 3 Pro is available in preview at $2 per million input tokens and $12 per million output tokens, for requests of up to 200,000 tokens. Google AI Studio offers free access with usage limits, while Vertex AI is aimed at businesses and applies its own terms and limits.

For you, the most important change is that building software with AI is moving closer to describing a goal than writing every line of code. That does not remove the need to review, test, and secure the result, especially when agents can take actions on a system. What you should watch now is how far that autonomy reaches in real projects and when the new terminal tools will become available to everyone.