Gemini 3 Deep Think, Google DeepMind’s specialized reasoning mode, now tackles complex science, research, and engineering challenges with refined capabilities. Developed in collaboration with scientists, the upgraded model excels in domains with ambiguous data and lacks clear solutions. Google AI Ultra subscribers can access it via the Gemini app, while researchers and enterprises can join the Gemini API Early Access Program to test its potential.

The updated Gemini 3 Deep Think achieves 48.4% on Humanity’s Last Exam and 84.6% on ARC-AGI-2, surpassing previous benchmarks. It also scores 3455 Elo on Codeforces and gold-medal-level results on the 2025 International Math Olympiad. These advancements highlight its ability to handle rigorous academic tasks, from theoretical physics to competitive coding.

Beyond academic benchmarks, Gemini 3 Deep Think drives real-world applications. At Rutgers University, it identified a flaw in a physics paper overlooked by humans. The Wang Lab at Duke University used it to design semiconductor crystal growth methods, achieving precise thin-film targets. Additionally, it converts sketches into 3D-printable reality files, streamlining prototyping.

With the updated Gemini 3 Deep Think, you can turn a sketch into a 3D-printable reality. Google’s Early Access Program invites researchers to explore its engineering and scientific applications. This marks a pivotal step in bridging advanced AI with practical innovation, setting a new standard for specialized reasoning systems.