Google officially unveiled Gemini 3 on November 18 (US time), presenting it as the company’s most intelligent and secure AI model to date. The goal: help users and developers bring any idea to life.

The launch comes as Google’s generative AI products continue to gain widespread adoption. According to the tech giant, the AI Overviews feature on Google Search is now used by 2 billion users each month, while the Gemini app has surpassed 650 million monthly users.

More than 70% of Google Cloud customers are utilizing AI services, and 13 million developers have built products using Google’s AI models.

Each Gemini generation has built upon the last. Gemini 1 introduced breakthroughs in native multimodality - the ability to naturally process multiple data types such as text, images, and audio simultaneously - along with enhanced long-context handling to absorb more extensive information.

Gemini 2 laid the foundation for agentic capabilities: empowering the model to perform complex, multi-step tasks autonomously, instead of merely answering questions. It pushed the boundaries of logic and reasoning, offering support for complicated ideas and workflows.

Now, Gemini 3 merges and amplifies all of these features, with a primary focus on advanced reasoning. The model is designed to understand nuanced meanings and subtle cues, even identifying minor hints or dissecting overlapping layers of complex problems.

It also excels at grasping the context and intent behind prompts, providing satisfying answers with fewer inputs.

According to Demis Hassabis, CEO of Google DeepMind, and Koray Kavukcuoglu, CTO and Chief Architect of Google AI, Gemini 3 represents a significant step toward artificial general intelligence (AGI).

They describe Gemini 3 as Google’s most powerful model yet for agent-based programming and "vibe coding."

What makes Gemini 3 special?

Gemini 3 launches with an early release of Gemini 3 Pro - a version already outperforming Gemini 2.5 Pro across all major AI benchmarks.

In terms of reasoning, Gemini 3 Pro scored 1501 on the LMArena leaderboard, reflecting a PhD-level reasoning capability. It also leads in specialized evaluations like Humanity’s Last Exam (37.5%) and GPQA Diamond (91.9%).

The model is deeply integrated across Google’s ecosystem to assist users in three core areas: learning, building, and planning.

For learning, Gemini 3 expands the limits of multimodal reasoning, allowing users to study in ways tailored to their needs. The model supports up to 1 million input tokens - enabling comprehension of vast documents.

It can decipher and translate handwritten formulas into multiple languages, create personalized cookbooks, and convert long academic papers or video lectures into interactive flashcards and visual aids. It can also analyze sports videos, pinpoint weaknesses, and develop training plans.

In Google Search, the new AI Mode now utilizes Gemini 3 to offer dynamic generative experiences in the user interface. It presents rich visual layouts, interactive tools, and real-time simulations based on user queries.

This helps users explore complex topics - like how RNA polymerase works - through engaging, interactive learning.

Google also introduced Antigravity, a new agentic development platform that allows programmers to focus on goals rather than individual steps.

Antigravity turns AI from a passive tool into an active collaborator, enabling agents to plan and execute complex software workflows autonomously while even verifying their own code.

With its robust planning abilities, Gemini 3 can now automate intricate processes like organizing Gmail or creating travel itineraries.

Google announced that Gemini 3 has already been deployed across its entire ecosystem. The model is available in the Gemini app for users subscribed to Google AI Pro or Ultra via Search’s AI Mode.

Developers can access Gemini 3 via the Gemini API on AI Studio, Google Antigravity, and Gemini CLI. Businesses can utilize it through Vertex AI and Gemini Enterprise.

The company also claims that Gemini 3 is its safest AI model ever. It has undergone Google’s most rigorous safety evaluation to date.

The model demonstrates reduced tendencies to conform to user bias and offers stronger resistance to prompt injection attacks.

Du Lam