Google launches Gemini 3 with multimodal processing, stronger reasoning & new agentic IDE

Google launched the Gemini 3 model family, which it describes as its most advanced and factually accurate to date. Gemini 3 Pro is now available to all users to all users on day one in the Gemini app and in AI Mode within Google Search, with a rollout also underway for paid subscribers. A higher tier version, Gemini 3 Deepthink, will follow in the coming weeks for Google AI Ultra users once safety reviews are complete.

Gemini 3 Pro introduces native multimodal processing for text, images, and audio, enabling tasks like turning recipe photos into a cookbook or creating flashcards from video lectures. In Search’s AI Mode, it can generate images, tables, grids, and simulations, supported by an upgraded query fan out system that better interprets intent and surfaces relevant content. The model leads current benchmarks with a 37.4 score on Humanity’s Last Exam and first place on LMArena, and Google notes that responses aim to be more direct and less focused on flattery.

The model’s improved reasoning and long horizon planning support more complex workflows and enable Gemini Agent, an experimental feature that can organize emails, manage information, and book travel for AI Pro and Ultra subscribers in the US. Users can access Gemini 3 Pro in AI Mode by selecting the Thinking option in the model menu. Google also introduced Antigravity, a new Gemini powered coding interface that brings together a prompt window, command line, and browser view to support multi pane agentic coding, similar to Warp Terminal.