On the morning of November 19, 2025 (Beijing Time), Google officially unveiled its new Gemini 3 series models, billing them as the "most intelligent" and "factually most reliable" AI systems to date. All users can directly access the new flagship model, Gemini 3 Pro, via the Gemini app, and this model will also be integrated into the search subscription service simultaneously. The standout feature of Gemini 3 Pro is its "native multimodality," which empowers it to handle text, images, and audio simultaneously. In Google Search's AI Mode, Gemini 3 Pro delivers results in a more intuitive format, encompassing images, tables, grids, and even simulated content. Its reasoning and agent capabilities have undergone substantial upgrades, making it more proficient in tackling complex tasks and endowing it with stable long-term planning skills. Currently, it tops the charts on LMArena, a widely recognized evaluation platform. Gemini 3 Pro is now accessible to all users through the Gemini app. Google AI Pro and Ultra subscribers in the United States can also opt for the "Thinking" mode to utilize this model in Search's AI Mode.

In terms of performance, Gemini 3 Pro shines brightly. In the Humanity’s Last Exam test, it scored 37.5% (45.8% with tool utilization), significantly outperforming GPT-5.1's 26.5% and Claude 4.5's 13.7%. In the ultra-challenging MathArena Apex math test, GPT-5.1 managed only 1 point, whereas Gemini 3 Pro achieved a remarkable 23.4 points, showcasing a substantial lead. In the multimodal GUI positioning test ScreenSpot-Pro, it attained an accuracy rate of 72.7%, double that of Claude 4.5 (36.2%).

Moreover, Gemini 3 Pro marks revolutionary advancements in front-end development. It can swiftly generate an interactive 3D garden HTML file in just 15 seconds based on textual descriptions, supporting ES module imports. It can also directly create runnable front-end code from webpage screenshots, with layout fidelity reaching walkthrough acceptance standards. In a mere 2 minutes, it can craft a fully functional Web OS (inclusive of terminal/editor/game), transcending the limitations of "decorative UI."

Regarding interaction, Gemini 3 Pro introduces the "GenerativeUI" function, which automatically generates tailored interface layouts based on the query content. For instance, when a user inquires about quantum computing, the system might generate an interactive interface featuring conceptual explanations, dynamic charts, and links to pertinent papers. For the same question posed by different audiences, the system tailors the interface design accordingly. For example, when explaining the same concept to children and adults, it adopts a cuter presentation for children and a more concise and clear one for adults.

This release also encompasses an agent system named Gemini Agent, currently in the experimental phase. This system is capable of executing multi-step tasks and connecting to Google services such as Gmail, Google Calendar, and Reminders. In inbox management scenarios, the system can automatically filter emails, prioritize them, and draft replies. Travel planning is another application area where users simply need to provide a destination and approximate time, and the system will consult the calendar, search for flight and hotel options, and arrange the itinerary. However, this feature is presently exclusive to Google AI Ultra subscribers in the United States.

In terms of multimodal processing, Gemini 3 Pro is constructed on a sparse mixture of experts architecture, supporting text, image, audio, and video inputs. The model's context window spans 1 million tokens, enabling it to process longer documents or video content. Tests conducted by Mark Humphries, a professor of history at Wilfrid Laurier University in Canada, reveal that the model's character error rate in recognizing 18th-century handwritten manuscripts is 0.56%, a reduction of 50% to 70% compared to the previous version.

Google has also introduced an optimized version known as Gemini 3 Deep Think, specifically crafted for complex reasoning tasks. This mode is currently undergoing safety assessments and is slated to be made available to Google AI Ultra subscribers in the coming weeks. In Google Search's AI Mode, users can click the "thinking" tab to observe the reasoning process of this mode. Compared to the standard mode, the Deep Think mode conducts a more thorough analysis before generating a response.