Google AI Ultra Subscribers Get Gemini 2.5 Deep Think for Complex Problem Solving
Google has rolled out Gemini 2.5 Deep Think, its latest advanced AI model, to Google AI Ultra subscribers. This new tool is designed to assist with iterative development and design, scientific and mathematical discovery, and algorithmic development and coding.
Internal evaluations indicate that Gemini 2.5 Deep Think has achieved a Bronze-level performance on the 2025 International Mathematical Olympiad (IMO) benchmark. In addition to making it available to Ultra subscribers, Google is also granting a select group of mathematicians access to the full version of the Gemini 2.5 Deep Think model that participated in the IMO competition.
The latest release incorporates insights from early trusted testers and recent research breakthroughs, representing a significant improvement over the version showcased at Google I/O earlier this year. This model is a variation of the AI that recently achieved a gold-medal standard at this year’s IMO. While previous iterations required hours to reason through complex mathematical problems, Google states that the current model is faster and more practical for daily applications.
Google positions Deep Think as a powerful instrument for creative problem-solving. The company has provided access to the Deep Think model within the Gemini app to mathematicians, such as Michel van Garrel, to aid in testing mathematical conjectures. Concurrently, a small group of mathematicians and academics are receiving access to the official, IMO gold-winning version of Gemini 2.5 Deep Think.
Deep Think operates using "parallel thinking" techniques, akin to how humans explore various perspectives on a complex problem. Google explains that this approach enables Gemini to generate and consider multiple ideas simultaneously, often revising or combining them over time, before presenting the optimal solution. To facilitate this, Google has extended the model’s "thinking time" or inference time, allowing it more opportunity to explore possibilities and devise creative solutions. Furthermore, the company has developed novel reinforcement learning techniques that encourage the model to utilize extended reasoning paths, thereby enhancing its problem-solving capabilities over time.
According to Google, Deep Think is versatile across several use cases. It can assist users with challenges requiring creative thinking, strategic planning, and step-by-step improvements. For design and development, the model can enhance both the aesthetics and functionality of web-based tasks. In scientific and mathematical discovery, it can aid in formulating and exploring conjectures, and even reason through complex scientific literature. Given its purported excellence in tackling difficult coding problems, Deep Think is also expected to be valuable in algorithmic development and code optimization.