Deep Think AI launches in Gemini app for Google AI Ultra subscribers
Google has announced the rollout of Deep Think, an advanced problem-solving feature, within its Gemini app, exclusively available to Google AI Ultra subscribers. This new capability is designed to enhance Gemini’s ability to tackle complex challenges, drawing on techniques developed from cutting-edge AI research.
Deep Think represents a significant advancement over previous iterations of Gemini, incorporating feedback from early testers and recent breakthroughs. It is a variation of the model that achieved a gold-medal standard at the 2025 International Mathematical Olympiad (IMO). While the full IMO model required extensive processing time for highly complex problems, the version integrated into the Gemini app has been optimized for speed and daily usability, demonstrating a Bronze-level performance on the 2025 IMO benchmark in internal evaluations.
At its core, Deep Think extends Gemini’s “thinking time” by employing parallel processing techniques. Much like a human exploring multiple angles of a problem, this approach allows Gemini to generate and simultaneously consider numerous ideas, hypotheses, and potential solutions. It can revise or combine different concepts over time before arriving at the most optimal answer. This extended inference time, coupled with novel reinforcement learning techniques, encourages the model to develop more intuitive and effective problem-solving strategies.
This enhanced capability positions Deep Think as a powerful tool for a variety of demanding tasks:
Iterative Development and Design: Deep Think has shown impressive performance in tasks requiring step-by-step construction and refinement, such as improving the aesthetics and functionality of web development projects.
Scientific and Mathematical Discovery: Its ability to reason through highly intricate problems makes it valuable for researchers. It can assist in formulating and exploring mathematical conjectures or navigating complex scientific literature, potentially accelerating the pace of discovery.
Algorithmic Development and Code: The feature particularly excels at challenging coding problems where precise problem formulation, careful consideration of trade-offs, and understanding of time complexity are crucial.
Deep Think’s performance is further validated by its state-of-the-art results on challenging benchmarks. When compared to other models without tool integration, Gemini 2.5 Deep Think achieves top performance across LiveCodeBench V6, which measures competitive coding prowess, and Humanity’s Last Exam, a benchmark assessing expertise across various domains including science and mathematics.
In parallel with its public release, Google is also providing a select group of mathematicians and academics with access to the official Gemini 2.5 Deep Think model that achieved the gold-medal standard at the IMO. This initiative aims to gather feedback on how the advanced model can enhance their research and inquiry, informing future improvements.
Google emphasizes its ongoing commitment to responsible AI development. Safety and ethical considerations are integrated throughout Deep Think’s training and deployment lifecycle. Testing has shown improved content safety and tone objectivity compared to Gemini 2.5 Pro, though it exhibited a higher tendency to refuse benign requests. As AI problem-solving abilities advance, Google is conducting deeper analyses of risks associated with increased complexity, including frontier safety evaluations and planned mitigations for critical capability levels. Further details on the safety outcomes of Gemini 2.5 Deep Think are available in its dedicated model card.
Google AI Ultra subscribers can access Deep Think today within the Gemini app. Users can enable the feature by toggling “Deep Think” in the prompt bar when selecting the 2.5 Pro model. It seamlessly integrates with existing tools like code execution and Google Search, and is capable of generating significantly longer responses. Looking ahead, Google plans to release Deep Think, both with and without integrated tools, to a set of trusted testers via the Gemini API in the coming weeks, aiming to explore its usability for developer and enterprise applications.
This launch represents a significant stride in Google’s mission to build more helpful and capable AI, reinforcing its commitment to using Gemini to push the frontiers of human knowledge.