Google's new Gemini 3 model arrives in AI Mode and the Gemini app (www.engadget.com)

Google has unveiled Gemini 3 Pro, its latest large model claiming improved reasoning, multi-modal understanding and “class-leading” vibe coding performance. Benchmarks show significant gains: a new top score of 37.5% on the difficult Humanity’s Last Exam (12.1 points ahead of Grok 4) achieved without web-search tools, and a 1,501 score on LMArena. The model is available now across Google products — in the Gemini app (pick “Thinking”), in Search’s AI Mode and AI Overviews (initially for AI Pro/Ultra subscribers), and to developers via the Gemini API, AI Studio and Vertex AI. Practically, Gemini 3 Pro powers a new Gemini Agent (building on Project Mariner) that can perform tasks on users’ behalf — e.g., manage emails if granted app access — and generates richer, interactive answers in Search (like embedded calculators). Google says the model augments AI Mode’s fan-out search strategy by issuing more targeted queries and surfacing content older models missed. For enterprises and devs, Google also released Antigravity, an agentic coding app that programs autonomously, and introduced Gemini 3 Deep Think (enhanced reasoning) for safety testing before broader release. The announcement signals a push toward more capable, tool-enabled LLMs and agent workflows, with attendant implications for productivity, search relevance and safety oversight.