GOOGLE Google's Gemini 3, launched November 18, represents a seven-month leap from Gemini 2.5,
achieving 37.4 on Humanity's Last Exam (18% ahead of GPT-5 Pro's 31.64) and 1501
Elo on LMSYS Arena. It supports a 1-million token context window for handling complex, layered
problems in science, math, and creative ideation.
Revolutionary Generative Interfaces: Unlike traditional text outputs, Gemini 3 dynamically assembles visual, interactive
UIs—e.g., creating a "Van Gogh Gallery" with contextual bios and images for each painting, resembling a digital magazine. This "air-reading"
(contextual empathy) reduces the need for detailed prompts, delivering "genuine insight" over flattery.
Gemini Agent: An experimental multi-step task handler integrates with Google services like Calendar, Gmail, and Reminders,
enabling autonomous workflows such as booking travel or generating reports. Deep Think Mode, a premium variant, boosts reasoning
to 41.0% on Humanity's Last Exam and 93.8% on GPQA Diamond for expert-level questions.
Availability: Rolling out immediately in the Gemini app, AI Mode (conversational search), and Google Search for Pro/Ultra
subscribers in the US, achieving 72% accuracy on benchmarks. Free access extended to U.S. college students via
Google AI Pro (includes 2TB storage). Enterprise rollout via Vertex AI and Google Workspace starts Q4 2025.
Real-World Applications: On November 17, Google expanded its AI-powered "Flight Deals" tool globally in Search, adding agentic
booking for reservations and events via AI Mode. This demonstrates Gemini 3's practical integration across Google's ecosystem for autonomous task handling.
Safety Focus: Gemini 3 underwent Google's most rigorous evaluations, reducing sycophancy, prompt injections, and cyber misuse
risks—addressing past criticisms of earlier Gemini versions and positioning it as enterprise-ready from day one.