Google quietly went live with the Gemini 3 Pro large model at 3am without a launch event. The model topped LMArena with 1501 Elo points, Human Last Exam (HLE) with 45.8%, MMMU-Pro up to 81%, Video-MMMU up to 87.6%, and outperforms the GPT-5.1. Its 1 million token context window supports long content processing, and its Deep Thinking capability hit a new high of 45.1% in the ARC- AGI-2 test hit a new high of 45.1%, and launched the Google Antigravity intelligent body platform. Users can experience it through the Gemini app or Google AI Studio.
- This summary is generated by AI analyzing the content of the article for informational purposes only.
At 3 a.m., the tech world was abuzz.
Without holding any conference, without warming up in advance, and without even announcing it on its official blog, Google quietly went live with the most anticipated big model of the year, the Gemini 3 Pro. This AI behemoth, which made its debut in the darkness, announced to the world with a set of breathtaking benchmarking data that the era of Google's AI has truly come to The AI giant has used a set of breathtaking benchmark data to announce to the world that Google's AI era is really here.
What's more surprising is that OpenAI CEO Altman personally tweeted his congratulations, and Musk also gave a rare nod of approval. You know, just a few hours ago, xAI just released Grok 4.1, this dark war between AI giants is more intense than we thought.
Performance Blitz: Comprehensively Crushing the GPT-5.1
The release of the Gemini 3 Pro is like dropping a nuclear bomb on a calm AI lake.
On the LMArena leaderboard, it topped the cliff with an Elo score of 1501, leaving all competitors behind. Most shockingly, it took a high score of 45.81 TP3T on this test, the Human Final Exam (HLE), demonstrating reasoning abilities approaching the level of a human PhD.
In terms of multimodal understanding, Gemini 3 Pro is also flawless. Faced with the complex task of mixing video, image, and text, it scored 811 TP3T in the MMMU-Pro test, and 87.61 TP3T in Video-MMMU, refreshing the industry perception.
Deep thinking: not just fast, but deep
What's truly amazing about the Gemini 3 series is its ability to think deeply. the Gemini 3, in Deep Think mode, hit an all-time high of 45.11 TP3T on the ARC-AGI-2 test. This test simulates a human's ability to solve problems when faced with complete unknowns and is a key indicator of AI's path to generalized intelligence.
One researcher told me that when he asked Gemini 3 to explain the principles of plasma flow in a tokamak device, the model not only generated an accurate 3D visualization code, but also created a poem that captured the essence of fusion physics. This depth of cross-disciplinary understanding and creative expression is a level of sophistication never before achieved by a model.
Millions of tokens: true omnimodal AI
Gemini 3's 1 million token context window completely frees up AI's ability to process long content. Imagine feeding an entire textbook, hours of lecture videos, and hundreds of pages of research papers all at once, and Gemini 3 understands the connections and context.
This has led to unprecedented application scenarios:
1. Translation of family handwritten recipes into multilingual digital recipes
2. Analyze your exercise videos to create a personalized training plan
3. Transforming complex academic papers into interactive learning guides
What is even more shocking is that this model is completely trained on the Google TPU. This bottom line of full-stack self-research is the true embodiment of Google's moat.
Ambient Programming: One Word, One Application
In the realm of programming, Gemini 3 has demonstrated near-magical abilities. It has been called the most powerful ambient programming and intelligent body coding model in Google's history, topping the WebDev Arena charts with 1,487 points.
In practice, when I asked it to create a 3D spaceship game, it not only perfectly realized the visuals, but also added interactivity and sound effects.
The most striking case is that a developer asked Gemini 3 to build a Game Boy emulator, and it not only implemented the functionality, but also drew the appearance of the Game Boy in SVG. This ability to understand complex requirements and execute them with precision is redefining the boundaries of human-computer collaboration.
The smart body revolution: from tool to partner
The release of Gemini 3 marks a key step for Google in the field of intelligences. It demonstrated unprecedented long-range planning capabilities in the Vending-Bench 2 test, which simulates the annual business of running a vending machine and tests the consistency of AI's decision-making in complex environments.
Gemini 3 Pro achieves a higher return on investment by maintaining consistent tool usage and decision making. This is not just a technological breakthrough, but a philosophical shift: from passive responders, AI has become active planners.
To support this shift, Google is simultaneously launching Google Antigravity, a revolutionary development platform for intelligences. this platform allows developers to collaborate with AI at a higher, task-oriented level, where intelligences have direct access to editors, terminals, and browsers to autonomously plan and execute complex, end-to-end software tasks.
How to experience
The Gemini 3 Pro preview is now fully live:
- Regular users can experience it through the Gemini app
- The AI model in search has been plugged into the new model
- Developers can access via Google AI Studio, Vertex AI
- Enterprise customers can use the Google Cloud platform on the
The much-anticipated Deep Think mode is expected to be available to Google AI Ultra subscribers in a few weeks.
If you want to use GPT Plus, Claude Pro, Grok Super, Sora official paid exclusive account, you will not be able to recharge it yourself can contact our professional team (wx: f15303420735)
Comment List (24):
Load More Comments Loading...