Google officially launches “Gemini 3”! What are the highlights of becoming the world’s smartest AI model?

Google announced the official launch of its latest generation large language model "Gemini 3", which is officially positioned as the smartest Gemini ever.
(Preliminary summary: Grok4 intelligence score beats OpenAI o3 and Gemini 2.5 Pro! Diversified competition turns to politics?)
(Background supplement: Gemini 3 is coming? There are frequent rumors on An announcement was made in the early hours of the morning, announcing the official launch of its latest generation large language model "Gemini 3". This is the third generation flagship AI model launched by Google in just two years following Gemini 1 at the end of last year and Gemini 2 at the beginning of this year. It has also been long-awaited by the market.

Google executives: It has evolved from "understanding text" to "understanding context"

Google CEO Sundar Pichai stated in the announcement that the Gemini series has accumulated more than 650 million monthly active users, 2 billion people have used AI Overviews, and has been adopted by 13 million developers. He emphasized that Gemini 3 is not only an upgrade in parameters and speed, but also represents a key transition for AI from passive responses to "active understanding of intentions and situations."

At the same time, Demis Hassabis, CEO of Google DeepMind, also bluntly said: "Gemini 3 is currently the model with the strongest multi-modal understanding ability in the world, and it is also our most powerful agent model."

Overview of the main progress of Gemini 3

According to the official technical report and immediately launched products, the current highlights of Gemini 3 include:

1. Significant improvement in reasoning ability

• LMArena ranking exceeded 1501 Elo, ranking first
• GPQA Diamond reached 91.9%, Humanity’s Last Exam 37.5% (without tools)
• Mathematical puzzle MathArena Apex set a new record of 23.4%
• After the launch of “Deep Think” mode, GPQA Diamond further improved to 93.8%

2. Leading the way in multi-modal understanding

• MMMU-Pro 81%, Video-MMMU 87.6%
• Can process text, images, videos, audio and code at one time, with a context length of 1 million tokens

3. Comprehensive upgrade of agent and developer experience

• SWE-bench Verified 76.2%, WebDev Arena 1487 Elo
• New agent-first development platform "Google Antigravity" is launched simultaneously, allowing AI agents to directly operate editors, terminals and browsers
• Supports Cursor, GitHub, Replit Instant access to third-party tools

4. Daily application scenarios are greatly expanded

• Family handwritten recipes can be translated and organized into electronic recipe books
• A training plan is automatically generated after analyzing sports game videos
• The "AI Mode" in the search is now equipped with Gemini 3. Provide real-time interactive visualization and simulation tools

Going online in phases from now on

• General users: Gemini App and AI Mode in search are available immediately (Pro/Ultra subscribers are preferred)
• Developers: Google AI Studio, Vertex AI, Google Antigravity have opened the preview version of Gemini 3 Pro
• Enterprise users: Vertex AI and Gemini Enterprise Simultaneously launched
• After the "Gemini 3 Deep Think" mode completes the security assessment, it will be open to Google AI Ultra subscribers within a few weeks

Google emphasized that Gemini 3 has completed the most stringent security testing in history and cooperated with institutions such as the British AISI to conduct third-party assessments. This release is regarded as Google’s last major AI update before the end of 2025, and also brings competition with top models from OpenAI, Anthropic, xAI and other manufacturers into a new round of fierce competition.