Google has begun rolling out the new Gemini 3 Deep Think mode to Google AI Ultra subscribers through the Gemini app. The feature, built to tackle advanced maths, science and logic-heavy queries, offers more powerful reasoning capabilities. Deep Think was first showcased last month with the launch of the Gemini 3 models and is now reaching users.
Google says the Gemini 3 Deep Think mode builds on the earlier Gemini 2.5 Deep Think versions, using advanced parallel reasoning to evaluate multiple hypotheses at once. This allows the AI to solve complex equations, multi-step logic problems and tasks that require deeper contextual understanding.
Google also noted that Ultra subscribers can start using Gemini 3 Deep Think by selecting “Deep Think” in the prompt bar and choosing Gemini 3 Pro in the model selector. According to Google’s blog, Deep Think shows improvements in key benchmarks like:
Humanity’s Last Exam: 41.0% (without tools)
ARC-AGI-2: 45.1% (with code execution)
Google’s Gemini 3 AI model, launched last month, delivers a major upgrade across the Gemini app, Search’s AI Mode and other Google services. The company says the new model offers more accurate, intelligent and natural responses than its predecessors. With enhanced reasoning and multimodal capabilities, Gemini 3 is better equipped to understand user intent, manage complex tasks and interpret various types of content.
Key improvements include:
More precise understanding of user intent
Better handling of long or detailed instructions
Ability to break down complex tasks with fewer follow-up prompts
Enhanced multimodal processing of text, images, screenshots, PDFs and videos
Improved document summarisation and image/graph analysis
Gemini 3 comes with larger context windows and better planning skills, allowing it to handle complex tasks like data organisation, detailed planning and research with improved accuracy. Google also says it’s their most secure model yet, offering stronger protection against deceptive prompts and sensitive or high-risk queries.