Google officially launched Gemini 3: currently the most powerful AI Agentic and Vibe Coding large language model.

Google today officially announced the launch of the new generation large language model Gemini 3, and simultaneously introduced Gemini 3 Pro across multiple services including Gemini App, SearchAI mode, AI Studio, and Vertex AI. Google states that Gemini 3 is currently the strongest multimodal and reasoning model, significantly outperforming its predecessor in various important AI benchmark tests such as science, mathematics, visual understanding, and long-term planning. In addition to Vibe Coding, Gemini 3 is also the strongest Agentic model, meaning that the AI can proactively assist you in completing tasks.

Google CEO: Gemini's understanding has evolved from reading text and images to being able to “read the air”.

Google CEO Sundar Pichai pointed out that since the launch of the Gemini series nearly two years ago, AI products have reached two billion users worldwide. Each generation of Gemini builds on the achievements of the previous one, with Gemini 1 making breakthroughs in native multimodality and long-context capabilities, allowing it to handle more and more complex information. Gemini 2 lays the foundation for agent ( agentic ) capabilities, pushing the limits of reasoning and thinking.

And now, Gemini 3: our smartest generation model, gathers all the capabilities of the Gemini series, allowing you to bring any idea to reality. It achieves SOTA (state-of-the-art) in reasoning, mastering depth and nuance. Whether capturing subtle cues in creative ideas or unraveling complex problems with layers of depth.

Gemini 3 can also better understand the context and intention behind your requests, allowing you to get the answers you truly need without struggling to come up with prompts. Astonishingly, in just two years, AI has evolved from being able to read text and images to being able to read the room.

Breakthrough in reasoning ability: Gemini 3 wins first place in LMArena, scientific reasoning, and mathematics tests.

Gemini 3 Pro sets multiple records in the latest reviews:

LMArena Leaderboard: Achieved 1501 Elo score.

Humanity’s Last Exam (Academic Reasoning): 37.5% (No Tools).

GPQA Diamond (Scientific Reasoning): 91.9%.

MathArena Apex (High Difficulty Math): 23.4%.

MMMU-Pro: 81%

Video-MMMU: 87.6%

SimpleQA Verified: 72.1% (Improvement in factual accuracy)

These results demonstrate that Gemini 3 Pro possesses high reliability in science, mathematics, and multimodal reasoning, capable of handling extremely complex problems.

Google synchronously released the Gemini 3 Deep Think reasoning model, achieving 45.1% in ARC-AGI-2 (including program execution), further enhancing its reasoning capabilities. Other highlights include:

Humanity’s Last Exam: 41.0%

GPQA Diamond: 93.8%

Gemini 3: Stronger Learning, Execution, and Planning Capabilities

Gemini 3 is currently the strongest vibe coding and agent-based programming model, with specific ratings including:

WebDev Arena: 1487 Elo (highest)

Terminal-Bench 2.0: 54.2% (Tool Operating Ability)

SWE-bench Verified: 76.2% (Large Program Tasks)

It also supports Google's brand new Google Antigravity proxy-based development platform, where AI can autonomously plan, write code, operate terminals, validate programs, and manage multi-step tasks in a variety of scenarios. Agentic AI refers to AI systems that can proactively take actions, plan multi-step tasks, and operate tools autonomously. The core concept is: AI is no longer just about answering but can actively complete tasks like an assistant.

For example, when I input: “Help me fetch today's ETH price and update Google Sheet.”, Agentic AI will check the API by itself and update Google Sheet.

Large language models mean that users input the same content, but due to the model's computations, they can have entirely different outputs. Gemini 3 can maintain consistent decision-making for a whole year in Vending-Bench 2, which means Gemini 3 can assist you in completing:

Book Local Service

Organize Gmail

Handling multi-step workflows

Gemini Agent is now open to Google AI Ultra users starting today. Google also stated that Gemini 3 is currently the most security-reviewed model, enhancing resistance to “flattery-style generation,” preventing Prompt Injection and online attacks. The Deep Think mode will be made available to Google AI Ultra subscribers after completing additional security testing.

This article announces that Google has officially launched Gemini 3: currently the most powerful AI Agentic and Vibe Coding large language model, first appearing in Chain News ABMedia.