Google officially unveiled "Gemini 3" – the newest and most powerful generation in the Gemini family, described as the company's "most intelligent model" to date. The model was developed by the Google DeepMind team and combines all the key capabilities of previous versions – from text and code to images, audio, and video – with a focus on deeper reasoning and a more accurate understanding of the user's intentions.
"Gemini 3" launches with the "Gemini 3 Pro" version, which is now available in the Gemini app, in AI Mode in Google Search, in AI Studio and Vertex AI, as well as in the new agentic platform for developers "Google Antigravity". Google emphasizes that for the first time a new model from the Gemini series is integrated into the search engine on the day of the premiere, to provide richer generative interfaces, dynamic visual layouts and interactive tools in real time.
According to the company, "Gemini 3 Pro" surpasses the previous "Gemini 2.5 Pro" in all leading AI benchmarks and is defined as the world's best model for multimodal understanding – from complex text tasks to combined analysis of text, images, and video. In 19 out of 20 tested large-scale benchmarks, Gemini 3 outperforms competitive models, including the top offerings of OpenAI.
For users, "Gemini 3" means more accurate and "less flattering" answers, with a clearer focus on usefulness and argumentation. The model is better at so-called "state-of-the-art reasoning" – unraveling complex tasks step by step, long-term planning, and analysis of "complex requests" with multiple conditions. For creating and debugging code, "Gemini 3" offers "vibe coding" and a deeper context when working with large code bases.
In parallel with the Pro version, Google announced "Gemini 3 Deep Think" – a mode with an even stronger focus on logical reasoning and complex multi-stage tasks. "Deep Think" goes through additional safety checks and is first provided to safety testers before becoming available to "Google AI Ultra" subscribers in the coming weeks.
For businesses and developers, "Gemini 3" is already available through Gemini Enterprise and Vertex AI – with full support for the analysis of text, images, video and files "in one step". Test partners such as Rakuten report over 50% better accuracy compared to basic models in transcribing multi-hour multilingual meetings, speaker recognition, and extracting structured data from poor-quality document photos.
The presentation of "Gemini 3" clearly shows Google's strategic goal – to offer a universal, multimodal, and "agentic" AI layer that will drive both the search engine and user applications, as well as corporate solutions and the development of complex agents. Additional Gemini 3 series models are expected in the coming months, and the company invites users and businesses to share feedback on the real-world scenarios in which the new model is changing everyday work.