Guotai Haitong: Alphabet Inc. Class C (GOOGL.US) Gemini 3 achieves disruptive leadership, accelerating the reorganization of the large model competition landscape.

Guotai Haitong released a research report stating that the release of Gemini 3 by Alphabet Inc. Class C (GOOGL.US) marks a new leap in large model technology, achieving breakthrough leadership in core capabilities such as inference, multimodal, and code generation. It innovatively launched the generative UI and intelligent agent platform Antigravity. This breakthrough validates the continued effectiveness of the Scaling Law and will accelerate the maturity of the AI application ecosystem. Guotai Haitong's main points are as follows: On Tuesday North American time, Alphabet Inc. Class C officially launched the new generation large model Gemini 3, achieving multidimensional breakthrough leadership. The new model has made significant progress in core reasoning capabilities, scoring 37.5% (without tools) in Humanity's Last Exam, up from 21.6% of 2.5 Pro; and surpassed GPT-5.1 (17.6%) by nearly double in the ARC-AGI-2, known as the "Turing test of the AI world," demonstrating close to human abstract reasoning abilities. In the multimodal understanding aspect, the new model has set new highs in complex scientific chart analysis and dynamic video understanding tests. Its excellent screen understanding capabilities lay a solid foundation for building truly practical AI agents. In the field of mathematical reasoning, the new model has evolved from only being able to handle basic operations to now being able to solve complex modeling and logical deduction problems, providing a reliable technical foundation for high-level applications such as engineering calculations and financial analysis. Gemini 3 has shown revolutionary progress in the fields of code generation and front-end design, not only completely reversing Alphabet Inc. Class C's competitive position in programming competitions, but also paving the way for commercialization through architectural innovation. It has achieved significant leading advantages on LiveCodeBench and ranked first in the Design Arena's website, game development, and four major competition areas. The groundbreaking aspect is that the model can not only generate functional code, but also possess "aesthetic intelligence," generating interactive interfaces that comply with modern design standards based on user intent, giving rise to a new paradigm of "generative UI." In terms of technical architecture, Gemini 3 adopts a new design of sparse MoE, supporting lengths of millions of tokens contexts, excelling in long document understanding and factual recall tests. Despite its API pricing being at the high end of the industry, the model provides a fine balance between performance and cost by increasing token efficiency and first-response accuracy, providing solid support for large-scale enterprise applications. Gemini 3 has made a qualitative leap in agent capabilities, becoming the first basic model to deeply integrate general Agent capabilities in consumer products. Its tool-use capabilities have improved by 30% compared to its predecessor, performing excellently in terminal environment tests and long-term business simulations, being able to autonomously plan and execute complex end-to-end tasks. Coupled with the newly launched Antigravity intelligent agent development platform, developers can perform task-oriented programming at higher abstraction levels, upgrading AI from an auxiliary tool to an "active partner." This breakthrough validates the continued effectiveness of the Scaling Law, driving the accelerating maturity of the AI application ecosystem and bringing about fundamental changes in the development paradigm of AI applications. Risk Warning: The iteration speed of large models is slower than expected, there is a shortage of computing power, and there are risks related to data privacy compliance.