Google has officially launched Deep Think for its Gemini app, making it available to Google AI Ultra subscribers. This advanced AI feature is designed to help users solve complex problems by extending the model’s ability to think more deeply and for a longer time before responding.
The rollout follows months of feedback from early testers and builds on research that helped a similar version of the model win a top award at the 2025 International Mathematical Olympiad (IMO), claims the tech giant.
While that original version was slow and aimed at detailed problem-solving, the version being released now is faster and designed for everyday use. According to Google’s own tests, it performs at a bronze-medal level on the latest IMO-style challenges.
Deep Think has been designed to assist in areas where creativity, planning, and step-by-step problem-solving are key. It is especially useful in fields such as science, mathematics, coding, and design. Google says that Deep Think does particularly well in building and refining complex systems, analysing tough scientific literature, and solving complicated coding tasks where time and resource planning are important.
The new system works by allowing the model more “thinking time” before giving an answer. This method lets Gemini look at multiple ideas at once, compare them, and revise its thoughts before reaching a final response. Reinforcement learning techniques help guide the AI along longer and more thoughtful reasoning paths.
In terms of performance, Deep Think has shown strong results on benchmarks such as LiveCodeBench V6, which tests code quality, and Humanity’s Last Exam, which measures deep knowledge and reasoning across various academic areas.
On the safety side, Google claims that Deep Think shows better content safety and a more neutral tone than earlier models, though it may sometimes refuse safe queries. The company says it is continuing to monitor and improve safety as the tool becomes more advanced.
To try Deep Think, AI Ultra subscribers can open the Gemini app, select the 2.5 Pro model, and turn on the Deep Think toggle. There is a daily limit on the number of prompts users can send. The model also works with tools like code execution and Google Search, and can generate longer responses than usual.
Google is planning to expand testing for developers and enterprise users soon via the Gemini API, with and without tool access. This will help assess how Deep Think performs in business and technical settings.