Gemini 3 is Google’s latest and most advanced AI model, integrated across their entire product ecosystem. Google launched three versions: Gemini 3, Gemini 3 Pro Preview, and Gemini 3 Deep Think. The model significantly outperforms previous frontier models in various benchmarks, including “Humanity’s Last Exam,” where it scored 37.5% without tools and 45.8% with code execution and search, far surpassing Gemini 2.5 Pro and competitors like Claude Sonnet 4.5 and GPT 5.1. It also excelled in the Vending Bench benchmark, which tests long-term planning and economic decision-making by managing a vending machine’s inventory, achieving a net worth of over $5,478 compared to the next best at $3,800.
Box.com conducted its own benchmark focusing on complex multi-step reasoning and document insight extraction, where Gemini 3 Pro showed a 22-point performance jump over Gemini 2.5 Pro, scoring 85% overall. The model demonstrated exceptional capabilities in industry-specific tasks such as healthcare, media, and financial services. This highlights Gemini 3’s strength in enterprise use cases, particularly in automating complex workflows involving unstructured data. Users can access Gemini 3 through Box AI Studio or the Box API, making it a powerful tool for businesses.
Gemini 3 Deep Think, a variant designed for extended reasoning with more tokens spent in the thinking phase, outperformed other models on reasoning and knowledge benchmarks, including scoring 41% on Humanity’s Last Exam and 93.8% on the scientific knowledge GPQA benchmark. It also showed a remarkable 10x improvement over Gemini 2.5 Pro in visual reasoning puzzles, a key test for generalized intelligence. The model supports multiple modalities—text, images, video, audio, and code—with a unique strength in video understanding, analyzing videos frame by frame rather than relying solely on transcripts.
One of Gemini 3’s standout features is its integration with Google Search in AI mode, where it dynamically generates user interfaces based on queries, creating custom search result pages. This innovation promises to transform how users interact with search results. Additionally, Google introduced Anti-Gravity, a new VS Code fork coding platform that supports Gemini models and other AI models, competing with existing AI coding tools. Gemini 3 also excels in long-horizon planning, as demonstrated by its performance in the Vending Bench 2 benchmark, maintaining and growing net worth over a simulated year.
Finally, Gemini 3 includes the Gemini Agent feature, allowing the AI to complete real tasks on users’ behalf, such as organizing emails with dynamic views and contextual email responses. Google also released a model card revealing that Gemini 3 is a brand-new foundation model, not a fine-tune of previous versions. It supports up to one million input tokens and 64,000 output tokens, running on Google’s custom TPU architecture for both training and inference. This combination of advanced capabilities and deep integration positions Gemini 3 as a leading AI model with broad applications across industries and everyday tasks.
