Google has officially rolled out Deep Think, its latest advanced AI feature, for Gemini app users subscribed to the Google AI Ultra tier. This marks a significant step forward in the tech giant’s efforts to provide cutting-edge tools for problem-solving and reasoning through its Gemini platform.
Deep Think is a specialized version of Google’s Gemini 2.5 model, fine-tuned for high-level reasoning, particularly in mathematical domains. Google has also begun providing select mathematicians and academics access to the full version of this model—one that recently earned gold-medal recognition at the International Mathematical Olympiad (IMO).
The latest iteration of Deep Think builds upon feedback from early testers and incorporates new research breakthroughs. Compared to the version showcased at Google I/O 2025, the current model demonstrates marked improvements in both speed and usability. While earlier versions took hours to reason through complex problems, Google claims this new model is significantly more responsive, enabling day-to-day use alongside deep research tasks.
According to Google, Deep Think now achieves Bronze-level performance on the 2025 IMO benchmark, a major milestone in aligning large language models with high-level academic and mathematical standards. It is also being tested by figures such as Michel van Garrel, a well-known mathematician, to explore mathematical conjectures and experimental problem solving.
A key enhancement in Deep Think is the extended inference time—often referred to as the model’s “thinking time”—which allows it to explore multiple solutions before arriving at a final response. This mimics the way human experts take time to contemplate and refine their ideas. Additionally, Google has integrated novel reinforcement learning techniques that guide the model along longer reasoning paths, resulting in richer and more accurate problem-solving capabilities over time.
This development underscores Google’s commitment to pushing the boundaries of AI-powered reasoning and creativity. With its selective rollout to AI Ultra subscribers and academic experts, the company is clearly targeting both high-end users and researchers who can test the limits of the model’s capabilities.
As Google continues to refine Deep Think, it may eventually extend the feature to broader audiences, potentially transforming the way users approach complex problem solving in education, research, and beyond.
