Google has unveiled Deep Think, a powerful new tool now available to users of its AI Ultra subscription via the Gemini app. Built on the same foundation that recently scored high in the prestigious International Mathematical Olympiad (IMO), this version has been optimized for speed and real-world usage while retaining its strong problem-solving abilities.

Unlike its slower, competition-grade predecessor, Deep Think in Gemini is designed to assist with day-to-day complex tasks. It utilizes a technique known as parallel thinking, enabling it to simultaneously explore multiple problem-solving routes. This makes it especially effective for analytical or creative tasks like debugging code, conducting scientific analysis, or structuring intricate design workflows.

In its launch statement, Google noted that Deep Think is “a major step up” from the original Gemini 2.5 model showcased earlier at Google I/O. The upgraded model reaches Bronze-level performance on the 2025 IMO evaluation, yet responds much faster for users needing quick, logical support.

CEO Sundar Pichai introduced the release on X (formerly Twitter), noting: “A version of Deep Think that reached gold-medal level at IMO is now live for Ultra subscribers in the Gemini app.” He also hinted that its best use cases include parsing scientific literature, solving algorithmic problems, and exploring research-level questions — all activities AI leader Demis Hassabis jokingly considers ideal for a Friday night.

To access the feature, users must toggle on Deep Think after selecting the 2.5 Pro model within the Gemini app. Initially, only a limited number of Deep Think prompts will be available each day. However, Google intends to gradually expand availability, including access via its API for selected researchers and developers.

Google emphasized that Deep Think isn’t just smarter — it’s safer, too. The company reports improvements in content filtering and bias reduction. However, the tool may lean toward caution, occasionally declining benign requests to ensure user safety.