Gemini 3 Pro is launching as a preview and will roll out across Google products, including the Gemini app, AI Studio, Vertex AI, and the AI mode in Google Search. It marks the first time a new Gemini model is available in Search on day one. Google says Gemini 3 delivers stronger contextual understanding and more nuanced responses. The answers are meant to be intelligent, concise, and direct, avoiding cliches and flattery in favor of real insight.
Stronger reasoning pushes Gemini 3 to new benchmark highs
Google highlights Gemini 3 Pro's performance with a long list of benchmark results. It reportedly leads the LMArena rankings with an Elo score of 1501 and shows what Google describes as PhD-level reasoning on tests like Humanity's Last Exam at 37.5 percent without tools and GPQA Diamond at 91.9 percent. That places it ahead of xAI's recent Grok 4.1 release. The model also posts strong scores in math at 23.4 percent on MathArena Apex and in multimodal understanding at 81 percent on MMMU-Pro.
Bild: Google
According to the official model card, Gemini 3 Pro is built on a sparse mixture-of-experts transformer architecture. Google trained it on a large multimodal dataset that includes publicly available web documents, licensed data, synthetic AI data, and user data from Google products and services. The model's knowledge cutoff is January 2025.
