August 2 News.GoogleIt was announced last night thatGemini The 2.5 Deep Think model is now live and available to Google AI Ultra subscribers in the Gemini app. The new model outperforms OpenAI's o3 and xAI's Grok 4 in several key benchmarks.
According to the foreign media Neowin reports, in May this year, Google in the Google I/O conference for the first time introduced the Gemini 2.5 Pro's Deep Think mode, using a new research method.Reasoning can be done by combining multiple hypotheses before answering the question. This release is an updated version of the feature, optimized with feedback from early internal beta users and the latest research. Google says the new version offers significant performance improvements over the initial I/O release.
Google also revealed that this model is a variant of a previous model (note: the aforementioned model has already won a gold medal at the International Mathematical Olympiad). In order to adapt it to everyday use scenarios, Google has optimized it for acceleration, and as a result, it will only reach the bronze level in the 2025 IMO benchmark test.
As you can see from the benchmark results, Gemini 2.5 Deep Think performs well in LiveCodeBench V6, Humanity's Last Exam, IMO 2025 and AIME 2025.
Currently, Google AI Ultra users can use Deep Think a limited number of times per day in the Gemini app, and when the Gemini 2.5 Pro model is enabled, the input field will show the "Deep Think" option. Google notes that this feature willAutomated integration of tools such as code execution and Google search. Google also plans to make versions of the Deep Think model with and without tools available to select developers through the Gemini API in the coming weeks.