Google is giving something new to its most advanced AI users to play, and it is called deep think. From today, Google AI Ultra can be subscribed to Ultra directly in the Gemini app. This new model is not just another repetition of Gemini 2.5; It is based on the same version that scored the top level score in the International Mathematical Olympiad (IMO), a global competition, where the best student brain goes to head-to-head on high-level mathematics problems. While that IMO version took hours to think through complex problems, one in the Mithun app has been tuned for speed and everyday purposes – although still capable of serious argument, according to Google’s internal test.

The focus with deep think is on better problems. It uses which is called parallel thinking, which means, instead of following a straight path for an answer, it considers several possibilities at once. This allows it to detect, compare, and even modify its approach. Google says it makes it particularly good for creative or layered tasks-things such as complex coding challenges, scientific research and step-by-step design projects.

Announcing the release, Google said, “This is a significant improvement that was first declared in I/O,” referring to Gemini 2.5 demo before that. The model now reaches bronze-level performance on the 2025 imo benchmark, while day-to-day works are far more responsible and experimentable.

CEO Sundar Pichai also shared the news on X (East Twitter), it was written: “We are bringing a version of deep thinking that in @GEMINIApp achieved gold medium position in IMO for ultra customers (+ official version is now in the hands of mathematicians). Friday night is considered a funny.”

For people eager about how to use it, the convenience is available through Gemini app on mobile. Customers need to enable the deep theme togle after selecting only 2.5 Pro models. At the time of launch, users indicate deep thinking in limited numbers each day. Over time, Google plans to expand its availability, including the Mithun API for selected developers and academics.

Beyond the performance, Google also states that it has worked on security. The new model shows better material filtering and fairness than previous versions, although it reportedly plays it safely and sometimes also refuses harmless questions.