Google has launched a major upgrade to Gemini 3 Deep Think, its advanced reasoning mode targeting complex, open-ended challenges in science, research, and engineering. In collaboration with the research community, Google emphasizes practical problem-solving capabilities of Deep Think, allowing it to reason effectively amid ambiguous rules and incomplete data. Highlighted by Sundar Pichai, the revised model has demonstrated impressive benchmark gains, scoring 48.4% on Humanity’s Last Exam and 84.6% on ARC-AGI-2. In competitive programming, it achieved an Elo rating of 3455 and gold medal-level performance at the 2025 International Math Olympiad. Additionally, Deep Think shows strong capabilities in physics and chemistry, posting high scores in related Olympiads. As a pivotal tool for researchers, Deep Think can analyze complex datasets and translate scientific knowledge into code. Currently accessible to Google AI Ultra subscribers, the system will be gradually made available via the Gemini API, based on user feedback.
Source link
