Google released Gemini 3, deploying it from day one across the Gemini app, Search, AI Studio, Vertex AI, the Gemini CLI, and its new Antigravity IDE — the first time its top-end model went into Search immediately. It posted a then-record 37.4 on Humanity's Last Exam, a test of general reasoning and expertise, became the first large language model to cross 1500 Elo on LMArena's human-preference leaderboard, and beat competitors on 19 of the 20 major benchmarks evaluated at launch, with 87.6% on Video-MMMU and 81% on MMMU-Pro. Like every Gemini since 1.5, it is a sparse Mixture-of-Experts model, which separates the model's total size from the computation spent on each token. A more compute-intensive Gemini 3 Deepthink was to follow for Google AI Ultra subscribers after further safety testing. Two years after Gemini 1.0, it marked Google leading on both capability and speed of deployment.