Google DeepMind announced the general availability of Deep Think, the company’s most advanced reasoning capability to date, now integrated within the Gemini app exclusively for Google AI Ultra subscribers. The release marks a significant milestone in DeepMind’s ongoing mission to develop more capable and helpful AI tools.
This latest iteration of Deep Think incorporates insights from trusted testers and leverages cutting-edge research advancements. Representing a substantial upgrade over the version introduced during Google I/O, the updated Deep Think model demonstrates notable improvements in both benchmark results and user feedback. It is a variation of the model that recently achieved gold-medal performance at the 2025 International Mathematical Olympiad (IMO). While that version required several hours to solve intricate mathematical problems, the newly released model offers faster response times suitable for daily use while still achieving Bronze-level performance on internal IMO benchmark tests.
As part of the launch, Google is also granting a select group of mathematicians and academics access to the gold-medal-level version of the Gemini 2.5 Deep Think model. “We look forward to hearing how it could enhance their research and inquiry, and we’ll use their feedback as we continue to improve this offering,” the team stated.
Unlocking New Possibilities with Parallel Reasoning
Deep Think builds on Gemini’s architecture by extending its parallel “thinking time.” Inspired by human problem-solving behavior where people consider multiple approaches before arriving at an answer Deep Think enables the model to simultaneously generate, revise, and synthesize a range of potential solutions. This extended inference period allows for deeper reasoning, fostering more creative and robust outputs.
To support this innovation, DeepMind has also implemented novel reinforcement learning techniques designed to enhance the model’s ability to follow extended reasoning paths. This makes Deep Think more intuitive and adaptive across diverse and complex problem domains.
Also Read: Google Introduces New AI Capabilities in Firebase and Flutter to Accelerate App Development
Driving State-of-the-Art Performance Across Fields
Deep Think has shown outstanding performance in scenarios that require strategic, creative, and iterative problem-solving. Notable applications include:
-
Iterative Development and Design: Deep Think has successfully demonstrated the ability to refine both aesthetic and functional aspects of complex web development projects.
-
Scientific and Mathematical Research: The model is capable of supporting researchers in formulating hypotheses, analyzing scientific literature, and accelerating mathematical discovery.
-
Advanced Algorithm Design: Deep Think excels at solving difficult coding challenges involving problem formulation, time complexity tradeoffs, and logical sequencing.
In performance evaluations, Gemini 2.5 Deep Think achieved state-of-the-art results on demanding benchmarks such as LiveCodeBench V6 and Humanity’s Last Exam, which test expertise across coding, science, and reasoning tasks.
A Responsible Approach to Innovation
DeepMind remains committed to building AI systems that are not only powerful but also safe and trustworthy. In internal evaluations, Deep Think demonstrated enhanced content safety and tone-objectivity when compared to Gemini 2.5 Pro, although it exhibited a slightly higher rate of refusals for benign queries. The company is continuing to assess emerging safety concerns tied to model complexity and is implementing mitigation strategies accordingly.
Additional details on Gemini 2.5 Deep Think’s safety and performance characteristics are available in the official model card.