Google rolls out Gemini 3 Deep Think mode with advanced reasoning power
Google has rolled out Gemini 3 Deep Think mode for AI Ultra subscribers in the Gemini app. The new mode boosts reasoning power using advanced parallel processing to handle complex math, science and logic tasks.
New Delhi: Google has started to implement Gemini 3 Deep Think mode on Google AI Ultra subscribers within the Gemini app. The new mode will be able to address higher reasoning science and logic problems that are complex in nature. It employs the use of advanced parallel reasoning and as such, the model can test two or more hypotheses simultaneously to achieve more accurate and consistent results.
Deep Think was already previewed when Gemini 3 Pro was launched last month. According to Google, the mode is providing a relevant jump in intelligence, particularly when it comes to jobs requiring a high level of analysis, multi-step reasoning and background. The feature is now available to the Ultra subscribers with the option of Deep Think on the prompt bar and then selecting Gemini 3 Pro on the app.
What’s new in Gemini 3 Deep Think
According to Google, Deep Think performs at the industry leading rate on a number of rigorous benchmarks. It has a 41.0% score on Humanity Last Exam (No tools) and 45.1% score on ARC-AGI-2 (Code execution). These will be based on the previous Gemini 2.5 Deep Think precedents, which have recently achieved gold medals in the major international events.
The new mode of reasoning also facilitates more developed code generation, enhancement of visualization and enhanced management of intricate problem solving tasks. According to Google, the system has iterative reasoning cycles that enable the model to get closer and more accurate results.
Broader improvements with Gemini 3
Gemini 3 was launched last month with significant new improvements in Google services, such as the AI mode of the Search platform and the Gemini application. The model has become more reasoning and it has better instruction handling and better understanding of the user intent. It is able to answer protracted detailed prompts with a lesser number of errors.
The model also has more powerful multimodal skills. It is capable of analyzing photos, screenshots, PDF files, and videos with greater precision, which allows performing actions, such as image analysis, document summarization and description of a graph or chart. According to Google, Gemini 3 has larger context windows as well, improved planning functionalities and additional safety measures, making it the safest model the company has published to date.