Google has launched the Deep Think mode for the Gemini 3 neural network. This special operating mode gives the model extra time for reasoning, enabling it to tackle complex mathematical, scientific, and logical problems that are typically out of reach for ordinary AI. The feature is already available to Google AI Ultra subscribers.

The key distinction of Deep Think is parallel reasoning. Unlike standard language models that generate responses sequentially, Deep Think processes multiple hypotheses simultaneously. This is similar to how humans try different approaches, discard dead-end logical paths, and arrive at the most well-founded conclusion.

This method is critical for tasks where a single misstep leads to a completely incorrect outcome and ensures unprecedented reasoning quality.

The results of Deep Think are impressive:

• On the ARC-AGI-2 abstract reasoning test, which evaluates logic and knowledge transfer, the model scored 45.1%. Most current AI systems do not surpass 20–30%.

• On the “Last Exam of Humanity” (a set of extremely challenging tasks), Gemini 3 Deep Think achieved a top score of 41.0%.

This new mode complements the standard Gemini for specific scenarios. It will be indispensable for scientists, technical university students, and developers — anyone who values the highest quality, accuracy, and well-founded reasoning over speed of generation.

ORIENT

Also read:

Tourism technologies: Dushanbe introduces the innovative “QR Guide” project

Success of Gemini 3 and Google’s Records: Larry Page Becomes the Second Richest Person in the World

AI Models Compete in Stock Trading: DeepSeek V3.1 Leads Ahead of GPT and Gemini