Google DeepMind has launched Gemini 3, its most advanced AI model to date, setting new benchmarks in reasoning, multimodal understanding, and coding capabilities. The model excels in complex tasks, from coding visualizations of plasma physics to translating handwritten recipes into multilingual cookbooks. Its Deep Think mode further enhances problem-solving, outperforming predecessors on benchmarks like Humanity’s Last Exam (41.0% without tools) and GPQA Diamond (93.8%).

Built on a foundation of state-of-the-art reasoning, Gemini 3 Pro scores 1501 Elo on LMArena and achieves 23.4% on MathArena Apex, surpassing earlier versions like Gemini 2.5 Pro. Its multimodal prowess shines with 81% on MMMU-Pro and 87.6% on Video-MMMU, enabling richer interactions across text, images, and video. The model also prioritizes factual accuracy, hitting 72.1% on SimpleQA Verified, reducing the need for excessive prompting.

Available immediately via the Gemini app, AI Studio, and Vertex AI, Gemini 3 integrates into Google’s ecosystem, including Search and the new agentic platform Google Antigravity. CEO Sundar Pichai highlighted its role in accelerating AI innovation, noting 2 billion monthly AI Overviews users and 13 million developers leveraging generative models. The release marks a strategic push to scale Gemini’s capabilities globally.

With a 1 million-token context window and leading multilingual performance, Gemini 3 redefines AI’s potential as a thought partner. It tackles niche challenges, like generating code for scientific visualizations or crafting poetry about fusion physics. As Deep Think mode expands to Ultra subscribers, Gemini 3 solidifies Google’s position in the AI arms race, blending technical rigor with creative problem-solving.