Google has announced a public version of its AI model, which won the International Mathematical Olympiad (IMO). This new version, called Gemini 2.5 Deep Think, is currently available exclusively to Google’s AI Ultra subscribers. Accessible for $250 per month, Google also offers a $125 discount for the first three months.
Gemini 2.5 Deep Think Announced
The new model stands out with its innovations in mathematical thinking. Deep Think utilizes an approach known as parallel thinking. This method allows the model to generate multiple solutions simultaneously, weighing them against each other to arrive at the optimal solution.
The model progresses by combining ideas when necessary, utilizing extended thinking periods. With the support of reinforcement learning techniques, its solution-oriented capabilities improve over time.
Google distinguishes this model not only for its intellectual capabilities but also for its productivity. Deep Think can provide longer and more comprehensive answers. It can also execute code when necessary and communicate with external tools like Google Search. This makes it particularly useful for multi-stage tasks and piecemeal development.
According to Google, this version demonstrates strong performance in areas such as scientific discovery, iterative design, software development, and mathematical analysis. The model’s scores on Humanity’s Last Exam, a multi-modal exam covering over 100 disciplines, support this strength.
Deep Think is positioned at a higher level not only in terms of capability but also in terms of content security. Google states that this model outperforms Gemini 2.5 Pro in neutral language and security criteria. Users can access this version through the Gemini app.
When the “Deep Think” option is enabled while 2.5 Pro is selected, the model is activated. Limited to a limited number of commands per day, it is poised to expand with API access in the future. Google plans to share both tool-supported and unsupported versions with trusted testers.
The current version of the model achieves bronze medal-level performance according to the criteria of the 2025 International Mathematical Olympiad. This is not the version that won the gold medal in the Olympics. This more advanced, high-level model, which can take hours to solve, is currently only available to academic institutions and special access groups.
Google recently publicly announced the success of the models it developed with OpenAI at the IMO. Now, the company has translated this success into a tangible product, offering users a tool capable of high-level mathematical and logical analysis with the Gemini 2.5 Deep Think.