Gemini 3: Google Reaches a New Milestone in Artificial Intelligence

Google has just unveiled Gemini 3, marking a major breakthrough in the development of generative artificial intelligence. After only two years of existence, the Gemini family now has more than 650 million monthly users, and this third generation promises to revolutionize the way we interact with AI. With cutting-edge reasoning, extended multimodal capabilities and new learning modes, Gemini 3 represents far more than a simple technical update.

Gemini 3 Pro: unprecedented reasoning that pushes the limits

The Gemini 3 Pro model now ranks at the top of the LMArena leaderboard with a score of 1501 Elo, surpassing all competing models. This spectacular performance is not limited to traditional benchmarks. On the formidable “Humanity’s Last Exam” test, Gemini 3 Pro reaches 37.5% without using tools, demonstrating reasoning equivalent to a doctoral level in complex scientific fields.

What really sets Gemini 3 Pro apart is its ability to grasp the depth and nuances of every interaction. Unlike previous generations, which tended to flatter the user, this new model adopts a more direct and insightful approach. It has a better understanding of the context and intent behind each request, which significantly reduces the need to rephrase queries. In just two years, AI has evolved from simply reading text to true situational understanding.

The model also excels at multimodal reasoning, combining text, images, video, and audio seamlessly. With scores of 81% on MMMU-Pro and 87.6% on Video-MMMU, Gemini 3 Pro sets a new benchmark for analyzing complex information across different formats. This versatility opens up fascinating possibilities, such as analyzing handwritten recipes in multiple languages or closely studying sports videos to improve one’s technique.

Gemini 3 Deep Think: when AI takes the time to think

Google is also introducing a specific mode called Gemini 3 Deep Think, designed for problems that require even deeper reasoning. This mode reaches 41% on “Humanity’s Last Exam” and 93.8% on GPQA Diamond, performances that even surpass the standard Gemini 3 Pro model. The most impressive result concerns the ARC-AGI benchmark, where Gemini 3 Deep Think achieves an unprecedented score of 45.1%, demonstrating its ability to solve entirely new challenges it has never encountered before.

This enhanced reasoning mode will be particularly useful for researchers, scientists, and anyone working on complex issues that require in-depth analysis. Google plans to make it available to Google AI Ultra subscribers in the coming weeks, after carrying out additional safety evaluations.

Learning with Gemini 3, a new way to learn, create, and plan

One of the most promising aspects of Gemini 3 lies in its practical applications for learning. The model is designed to synthesize information on any subject across multiple formats. You can submit long academic papers, full video lectures, or tutorials, and it will automatically generate interactive revision sheets, visualizations, or any other format suited to your learning style.

For developers, Gemini 3 is a revolution. It tops the WebDev Arena leaderboard with a score of 1487 Elo and achieves 54.2% on Terminal-Bench, demonstrating its ability to use tools to control a computer via a terminal. This excellence in “vibe coding” makes it possible to create rich, interactive web interfaces using natural language instructions, radically transforming the way applications are developed.

Google is also launching Antigravity, a new agentic development platform that leverages these advanced capabilities. Agents can now autonomously plan and execute complex end-to-end software tasks, validate their own code, and actively collaborate with developers rather than merely serving as simple assistant tools.

More autonomous AI for everyday tasks

Gemini 3 also excels in long-term planning. On the Vending-Bench 2 test, which simulates managing a vending machine business over a full year, Gemini 3 Pro maintains consistent tool use and generates higher returns than all competing models. This capability translates concretely into assistants that can handle complex processes such as booking local services or organizing a Gmail inbox.

Google AI Ultra subscribers can already test these agentic capabilities via Gemini Agent in the Gemini app. This feature will gradually be rolled out to other Google products over the coming months.

Availability and Outlook

Gemini 3 Pro is available starting today in the Gemini app for all users, as well as in the AI mode of Google Search for Pro and Ultra subscribers. Developers can access it via AI Studio, Vertex AI, and the new Antigravity platform, as well as on third-party platforms such as Cursor, GitHub, or Replit.

Google states that Gemini 3 is its most secure model to date, having undergone the most comprehensive series of security evaluations ever carried out for one of the company’s AI models. With increased resistance to prompt injection and better protection against cyberattacks, Google seems to be taking seriously the security issues surrounding the deployment of AI.

This third generation of Gemini marks a decisive step toward artificial general intelligence. By combining advanced reasoning, native multimodality, and agentic capabilities, Google offers a vision of AI that no longer merely answers questions. It becomes a true partner in thinking and action. It remains to be seen how users will adopt these new capabilities and what innovative applications will emerge from this ever-evolving ecosystem.

[1] A new era of intelligence with Gemini 3