Google Launches Gemini 3 With Major AI Upgrades in Reasoning and Multimodal Performance

Google has officially unveiled Gemini 3, the next generation of its flagship artificial intelligence model featuring significant improvements in reasoning, multimodal understanding, and autonomous task execution. The new model is now available across multiple Google products including Search, the Gemini app, and developer platforms.

Benchmark Performance Breakthrough

The Gemini 3 Pro model, released in preview, demonstrates superior performance compared to its predecessor Gemini 2.5 Pro across key industry benchmarks. Google reported exceptional scores on LMArena, Humanity’s Last Exam, GPQA Diamond, and MathArena Apex. The model also shows enhanced capabilities in multimodal assessments including MMMU-Pro and Video-MMMU, along with high accuracy ratings from SimpleQA Verified.

Advanced Reasoning with Deep Think Mode

Google introduced Gemini 3 Deep Think, specifically designed for complex reasoning tasks. Early testing reveals improved performance on challenging problem-solving benchmarks such as Humanity’s Last Exam and ARC-AGI 2. This advanced mode will become available to Google AI Ultra subscribers after completing additional safety reviews.

Search Integration and Developer Tools

For the first time, Google is integrating its latest AI model directly into Search at launch. The AI Mode in Search will leverage Gemini 3 to generate more dynamic layouts, interactive elements, and detailed visual explanations.

Developers can access the model through Gemini API, Google AI Studio, Vertex AI, Gemini CLI, and the new agent-focused development environment called Google Antigravity.

Antigravity: AI-Powered Development Platform

Google Antigravity represents a significant advancement in AI-assisted development, enabling programmers to use AI agents for planning, writing, and validating code within an integrated workspace. The platform incorporates Gemini 3 Pro along with Google’s latest computer-use model for browser-based actions and an updated image editing model.

Enhanced Long-Term Planning Capabilities

Gemini 3 demonstrates improved long-term planning abilities, achieving the highest score in the recent Vending-Bench 2 test that evaluates AI systems’ capacity for consistent decision-making across extended tasks. These capabilities are being made available to end users through Gemini Agent within the Gemini app.

Google confirmed that additional models in the Gemini 3 series will be released in the coming weeks, expanding the capabilities of their AI ecosystem even further.