Google DeepMind introduced Gemini 2.5 Deep Think — a new feature for the Gemini AI model, now available for Ultra subscription users in the Gemini app. This version of the model uses a so-called “parallel thinking approach,” allowing it to consider multiple ideas simultaneously, combine them, and find optimal solutions for complex tasks. The model can generate longer and more detailed responses and automatically uses tools such as code execution and Google Search.
Gemini 2.5 Deep Think demonstrates high results in tests on logical thinking, knowledge, and programming. In the LiveCodeBench 6 benchmark, it scored 87.6%, surpassing models from OpenAI and xAI, and in the Humanity’s Last Exam test, it achieved 34.8% without using additional tools. The model is particularly effective for tasks requiring creativity, strategic planning, and step-by-step improvement, especially in web development, scientific research, and complex mathematical problems.
For Ultra subscribers, the Deep Think feature is activated by a separate button in the Gemini app, and the number of daily requests is limited. The full version of the model, which won a gold medal at the International Mathematical Olympiad, is available only to select scientists and mathematicians, as it requires more time to solve complex problems. The public version of Deep Think is optimized for faster performance and everyday use, while maintaining a “bronze” level of performance by Olympiad standards.
Google plans to provide access to Gemini 2.5 Deep Think via API to a limited group of developers and enterprises for testing new usage scenarios. The company emphasizes that the model undergoes a multi-level security audit and has additional protective filters to prevent dangerous responses and misuse. The Ultra subscription costs $249.99 per month in the US, with a promotion for new users.
