Google has launched a powerful upgrade for its Gemini AI app called Deep Think, available exclusively to Ultra subscribers. This new feature is based on a high-level model that previously demonstrated Bronze-level performance in the International Mathematical Olympiad (IMO) — a competition known for its complex, abstract math problems.
Olympiad-level reasoning in your pocket
While the IMO version of the model used to take hours to work through problems, the version now available in the Gemini app is tuned for faster, practical use, according to Google. Despite the speed optimisation, it retains the deep reasoning ability needed for multi-step problem-solving, making it a valuable tool for coders, researchers, and creators.
What is parallel thinking?
Deep Think uses a method called parallel thinking — instead of pursuing a single path to reach an answer, the AI evaluates multiple possibilities simultaneously. This allows it to revise its approach, compare ideas, and handle complex coding tasks, scientific literature analysis, and other layered challenges with greater flexibility and insight.
How to use Deep Think
To access the feature:
- Open the Gemini app on mobile
- Select the 2.5 Pro model
- Toggle on Deep Think from the settings
Note: Daily usage is limited at launch, but Google plans to expand access through the Gemini API for developers and academic institutions soon.
Sundar Pichai and Google confirm advanced AI upgrade
Sundar Pichai, the CEO of Google has reportedly confirmed the rollout in a post on X (formerly Twitter), saying users now have access to a version of Deep Think that “achieved gold-medal status at IMO,” referring to its top-tier performance in internal testing.
He encouraged users to try it for anything from complex time complexity analysis to deep scientific reasoning, calling it perfect for "a fun Friday night" for AI enthusiasts.
Safety and reliability also improved
According to Google, the updated model has stronger safety filters and is more objective than its predecessors. However, it sometimes errs on the side of caution, refusing even harmless queries to avoid risks.
What’s next?
This is a significant improvement over the earlier Gemini 2.5 demo shown at Google I/O 2025, marking a step forward in accessible, high-level AI reasoning. The release of Deep Think highlights Google’s ambition to bring expert-level AI tools into the hands of everyday users.
