Google has unveiled its new Gemini 3 and Gemini 3 Pro artificial intelligence models, touting significant advancements in reasoning, multimodal understanding, and benchmark performance.
The Gemini 3 model introduces a “Thinking Mode” feature. This allows the AI to pause and process complex inquiries before generating a response. This capability aims to enhance performance in areas such as mathematical problem-solving, strategic planning, and programming tasks. Users can either let the system adapt its thinking level automatically or set it manually.
Gemini 3 Pro is engineered to interpret diverse data formats more effectively. Beyond traditional text, it can process and understand images, audio, and video concurrently. For instance, the AI can analyze sports footage to suggest proper techniques or decipher handwritten documents from a photograph. This marks a notable improvement over previous versions.
Both models have reportedly achieved top rankings in multiple global benchmark tests. The Gemini 3 Pro scored 37.5% on Humanity’s Last Exam (HLE), a rigorous evaluation. An even more advanced “Deep Think” version of Gemini 3 achieved a score of 41% on the HLE. Google states these results surpass competitors, including GPT-5.1.
Google also claims the new models demonstrate increased data accuracy. They are designed to reduce instances of “hallucination,” where the AI generates incorrect or fabricated information, compared to their predecessors.
A notable feature, “Vibe Coding,” allows users to create applications, games, or simple web pages. This functionality enables development simply by typing instructions, without requiring extensive coding knowledge. The AI then writes the necessary code and constructs the user interface for immediate use.
The Gemini 3 Pro model is currently available for public use. It can be accessed through the Gemini application and its dedicated website.
The more advanced Gemini 3 “Deep Think” variant will be released later. This version, with its higher analytical capacity, will initially be made available to Google AI Ultra subscribers and a select group of safety testers.