Google is starting to deploy a new artificial intelligence model, Gemini 3, which it calls the «most accurate» and «smartest» in the world and a big step towards AGI.
Google is starting to deploy a new artificial intelligence model, Gemini 3, which it calls the «most accurate» and «smartest» in the world and a big step towards AGI.
«This is the world’s best model for multimodal understanding and our most powerful model for agents and vibecoding to date, delivering richer visualization and deeper interactivity — all powered by cutting-edge reasoning methods», said the developers of Gemini 3.
Gemini 3 Pro’s multimodality means that the AI can process text, images, and audio simultaneously, rather than separately. For example, Gemini 3 Pro can be used to translate photos of recipes and then turn them into a cookbook, or to create interactive flashcards based on a series of video lectures.
Google has announced that it is starting to release the Gemini 3 Pro in a pre-release version. This model is designed for everyday use for learning, creating, and planning.
The updated AI model will also enable «generative interfaces,» a tool Google is testing in Gemini Labs that allows Gemini 3 Pro to create a magazine-style visual format with scrollable images or a dynamic layout with a customizable user interface.
Another new feature of the Gemini 3 Pro in AI Mode is Google’s AI-powered search feature, which can generate images, tables, grids, and simulations based on a query. The model is able to perform more searches using an updated version of Google’s «query branching technique,» which now not only breaks down questions into parts that it can search on the user’s behalf, but also better understands intent to help «find new content that might have previously been missed.»
Google has hinted that Gemini 3 Pro is less prone to flattery than OpenAI’s ChatGPT. According to the company, its AI provides «intelligent, concise, and direct responses that replace clichés and flattery with real understanding — telling you what you need to hear, not just what you want to hear.»
«Gemini 3 Pro leads the LMArena rankings with a groundbreaking Elo score of 1501. It demonstrates PhD-level reasoning, achieving the highest scores on Humanity’s Last Exam (37.5% without using any tools) and GPQA Diamond (91.9%). It also sets a new standard for advanced models in mathematics, reaching a new high of 23,4% on MathArena Apex,» Google boasts.
Gemini 3 Pro also has better reasoning and agency capabilities, allowing it to perform more complex tasks and «reliably plan for longer periods of time.» The AI model powers an experimental Gemini Agent feature that can perform tasks on behalf of the user in the Gemini app, such as viewing and organizing emails or finding and booking travel.
The company also introduced Gemini 3 Deep Think, an advanced reasoning mode that further enhances the performance of Gemini 3. Access to this model will first be opened to security testers before being made available to Google AI Ultra subscribers.
Gemini 3 Pro is available in the Gemini app for all users starting today, and Google AI Pro and Ultra subscribers in the US can try out Gemini Agent in the Gemini app, as well as Gemini 3 Pro in AI Mode by selecting «Thinking» from the model menu.
