- Google's most capable Gemini 2.5 Deep Think AI model is finally available for public users.
- It's a faster version of the AI model that achieved the Gold medal at IMO 2025.
- It's rolling out to Google AI Ultra subscribers in the Gemini app.
Google has finally released the Gemini 2.5 Deep Think mode, which has delivered breakthrough performance in key areas like mathematical research, coding, and scientific reasoning. This is a faster variation of the same AI model that recently achieved the Gold medal at the IMO 2025.
The IMO variant takes hours to reason on complex problems, so Google has developed a faster version for day-to-day usage, and it’s being released as Gemini 2.5 Deep Think. Even then, this version achieves the Bronze medal on the IMO 2025 benchmark, according to internal evaluation.
The new Gemini 2.5 Deep Think mode utilizes new methods such as parallel thinking to explore many ideas at once, extended thinking time to reason through complex problems, and novel reinforcement learning (RL) techniques to “make use of these extended reasoning paths.”
These are new research ideas in the AI field, and both Google DeepMind and OpenAI are working on parallel test-time compute. This approach is yielding great results in generating mathematical conjectures, creative problem solving, coding challenges, understanding scientific literature, and more.
In the Humanity’s Last Exam benchmark, Gemini 2.5 Deep Think scored the highest, achieving 34.8% and outranking OpenAI o3 and Grok 4. Similarly, in LiveCodeBench, Gemini 2.5 Deep Think does much better than frontier AI models from OpenAI and xAI.
Google says Gemini 2.5 Deep Think is rolling out to Google AI Ultra subscribers in the Gemini app, those who pay $250 per month. You can choose the Gemini 2.5 Pro model and select the “Deep Think” toggle to activate it. Meanwhile, Google is also releasing the full version of Gemini 2.5 Deep Think to select mathematicians.
