Gemini 3 Pro Release Notes

Google DeepMind has introduced Gemini 3 Pro, a high-intelligence model designed to enhance agentic workflows and complex zero-shot tasks. Gemini 3 Pro outperforms previous versions across all major AI benchmarks and specifically surpasses Gemini 2.5 Pro in coding and reasoning capabilities.

Agentic Coding and Development Tools

Gemini 3 Pro serves as a foundation for agentic coding, demonstrating a 54.2% score on Terminal-Bench 2.0, which measures a model's ability to operate a computer via a terminal. The model is integrated into several developer tools, including Android Studio, Gemini CLI, Cursor, GitHub, JetBrains, Manus, and Cline.

Google Antigravity

Google Antigravity is a new agentic development platform that allows developers to act as architects while managing autonomous agents across workspaces. These agents can operate across the editor, terminal, and browser to plan and execute complex software tasks, communicating progress via detailed artifacts. The platform is available in public preview for MacOS, Windows, and Linux.

Gemini API Enhancements

New tools have been added to the Gemini API to support agentic workflows:

Bash Tools: A client-side bash tool allows the model to propose shell commands for local filesystem navigation and system automation. A hosted server-side bash tool is also available for secure prototyping and multi-language code generation.
Structured Outputs: Grounding with Google Search and URL context can now be combined with structured outputs, facilitating the extraction of data into specific formats for downstream agentic tasks.

Vibe Coding and Rapid Prototyping

Gemini 3 Pro enables "vibe coding," where natural language is the primary syntax for application development. The model scored 1487 Elo on the WebDev Arena leaderboard, reflecting its ability to translate high-level ideas into interactive applications via single prompts.

Google AI Studio Build Mode

Google AI Studio's Build mode allows developers to go from a prompt to an AI-native app quickly by automatically wiring models and APIs. The model's improved instruction following allows for the creation of fully functional apps—such as retro games or interactive landing pages—from single prompts, voice notes, or napkin sketches.

Multimodal Understanding and Spatial Reasoning

Gemini 3 Pro is positioned as a leading model for complex multimodal understanding, setting new benchmarks on MMMU-Pro (image reasoning) and Video MMMU (video understanding). It features a 1 million-token context window.

Visual and Spatial Reasoning

Document Understanding: The model moves beyond OCR to provide intelligent reasoning over complex documents.
Spatial Understanding: Gemini 3 Pro excels in embodied reasoning tasks, including trajectory prediction and pointing, which are applicable to autonomous vehicles, robotics, and XR devices.
Screen Understanding: The model can interpret desktop and mobile OS screens, including user intent derived from mouse movements and screen annotations, to power computer-use agents.

Video Reasoning

Gemini 3 Pro supports high-frame-rate understanding to capture rapid actions and utilizes long-context recall to synthesize narratives across hours of footage. According to Jay Wu, CTO of OpusClip, the model provides a 32% increase in speed over previous implementations for video agent reasoning and tool calls.

Technical Specifications and Availability

Gemini 3 Pro is available in preview via the Gemini API in Google AI Studio and Vertex AI for enterprises.

Pricing for prompts 200k tokens or less:

Input: $2/million tokens
Output: $12/million tokens

API Controls: To support deeper reasoning, the API now includes a new thinking level parameter, granular media resolution settings for multimodal vision processing, and stricter validation for thought signatures to maintain context across multi-turn conversations.