Google Unveils Gemini 3 Deep Think – Powerful AI for Science
Credit: Google
Google has released a major update to Gemini 3 Deep Think, a platform powered by reasoning artificial intelligence designed to solve pressing problems in science and engineering.
Google cites practical examples of its use. For example, mathematician Lisa Carbone of Rutgers University used Deep Think to validate a technical research paper in a field with virtually no training data. The model was able to identify a subtle logical error that had previously escaped peer review. And at Duke University, AI helped optimize the creation of complex crystals to search for new semiconductors. The model developed a precise “recipe” for growing thin films larger than 100 microns—a task traditional methods had failed to achieve.
The updated Deep Think mode continues to push the boundaries of intelligence and reaches new heights according to rigorous academic criteria, Google writes. The company lists record-breaking results on several key benchmarks: 48.4% (unaided) on the “Humanity’s Final Exam,” 84.6% on ARC-AGI-2, an Elo rating of 3455 on Codeforces, and a gold medal score at the 2025 International Mathematical Olympiad.
Beyond mathematics and programming, Deep Think has also gained strength in other scientific fields. The model achieved gold-medal-level results in the written sections of the 2025 Physics and Chemistry Olympiads, and scored 50.5% on the CMT-Benchmark, demonstrating its mastery of advanced theoretical physics.
Engineering applications have also received attention. The updated Deep Think is capable of interpreting complex data and modeling physical systems using code. As a demonstration, Google describes a scenario in which the model transforms a sketch into a form suitable for 3D printing, including generating a file for fabricating the object.
