Google Releases Gemini 3 Pro, Its Most Powerful AI Model Yet
Google has released Gemini 3 Pro in public preview, offering free access via Google AI Studio. However, it is currently not available in the consumer Gemini app or to paying Gemini Pro subscribers. The model supports multiple input and output modalities including text, images, audio, and video, and can process up to 1,000,000 tokens of context, equivalent to about 700,000 words or roughly 10 novels.
Gemini 3 Pro demonstrates superior performance compared to its predecessor, Gemini 2.5 Pro, scoring 37.5% versus 21.6% on Humanity's Last Exam and 31.1% versus 4.9% on ARC-AGI-2 visual reasoning. It employs a sparse mixture-of-experts architecture, routing inputs to specialized subnetworks to reduce computational demands while maintaining performance.
The model was trained on diverse data including web documents, code, images, audio, video, and synthetic data, with filtering to exclude pornographic content, violence, and child safety violations. Training utilized Google TPUs with JAX and ML Pathways, enabling generation of outputs up to 64,000 tokens. Gemini 3 Pro's knowledge cutoff is January 2025, but it may occasionally hallucinate and can be slower in some scenarios. An official model card has not been released yet.
Availability extends across Google AI Studio, Vertex AI, and the Gemini API, although it is not accessible through the Gemini consumer app at this time. In the competitive landscape, Gemini aims to contend with models such as Claude Sonnet 4.5, Grok 4.1, and GPT-5.1. Google reports that Gemini has approximately 650 million monthly users compared to ChatGPT's 800 million weekly users.
