Google Launches Gemini 3 Deep Think for Complex Reasoning

Google launched Gemini 3 Deep Think, a specialized AI reasoning mode optimized for complex problem-solving in science, research, and engineering. The model is designed to handle challenges with incomplete or contradictory data where no single correct answer exists. It achieved gold-medal level performance in the 2025 International Math, Physics, and Chemistry Olympiads, recorded an Elo rating of 3455 on Codeforces, and scored 84.6% on the ARC-AGI-2 benchmark.

Real-world applications of the model include identifying a logical flaw in a physics paper at Rutgers University and optimizing semiconductor crystal growth at Duke University's Wang Lab. The system can also convert hand-drawn sketches into 3D-printable objects. Access is provided to Google AI Ultra subscribers via the Gemini app and to select enterprises and researchers through an API early access program.

Following the model's release, Google DeepMind introduced Aletheia, an AI agent based on Gemini 3 Deep Think. Aletheia analyzes scientific hypotheses and uses web tools to reduce the citation of nonexistent sources. While the agent proposed novel contributions to unsolved problems by mathematician Paul Erdős, developers noted that Aletheia remains more error-prone than human experts and occasionally simplifies complex problems to arrive at answers.