On November 19th, Google released the Gemini 3 model. Demis Hassabis, CEO of DeepMind, and Josh Woodward, head of the Gemini team, jointly gave an exclusive interview.
With this upgrade, Gemini 3 has gained the capability of Generative UI. When users inquire about Vincent van Gogh’s life, it can instantly build an interactive page containing images and a timeline; when complex calculations are involved, it can directly generate a customized mortgage calculator.
This leap from answering questions to building applications marks that large model applications are moving beyond simple dialogue boxes and entering the stage of dynamic software generation.
The reasoning ability of the Gemini 3 model has been significantly enhanced. According to Woodward, previous-generation models often lost their train of thought during the 5th or 6th step of reasoning, while Gemini 3 can maintain coherent logic for 10 to 15 steps in complex tax planning or long code debugging, greatly improving reliability when handling complex tasks.
In the interdisciplinary doctoral-level problem set exam known as “Humanity’s Last Exam”, the score of Gemini 3 Pro surged from 21.6% (achieved by its predecessor Gemini 2.5 Pro) to 37.5%, far exceeding GPT-5.1’s 26.5%. In the SimpleQA Verified test, Gemini 3 Pro achieved an accuracy rate of 72.1%, which is more than double that of GPT-5.1 and Claude Sonnet 4.5, significantly reducing the hallucination phenomenon commonly seen in previous models.
The most groundbreaking upgrade this time appears in the field of visual intelligence. Gemini 3 Pro scored a high 72.7% in the ScreenSpot-Pro test, which is specifically designed for screen understanding and UI interaction—nearly 20 times the performance of GPT-5.1. This means that AI Agents can distinguish between buttons, menus, text, and context, rather than just recognizing objects in images. It provides a foundation for AI Agents to achieve more advanced automated computer operations and greatly enhances the practicality of the model as a digital work agent.