Gemini 3 Deep Think is a significant upgrade designed to tackle complex challenges in science, research, and engineering. Developed in collaboration with researchers, this specialized reasoning mode aims to offer practical solutions where typical approaches struggle due to messy data or ambiguous problems. It's now available in the Gemini app for Google AI Ultra subscribers and through the Gemini API for select researchers and enterprises. Early users have demonstrated its capabilities in various fields. For instance, mathematician Lisa Carbone utilized Deep Think to identify a subtle logical flaw in a high-energy physics paper that eluded peer review. At Duke University's Wang Lab, the tool optimized crystal growth processes for semiconductor material discovery, successfully creating thin films over 100 μm. Anupam Pathak from Google also leveraged Deep Think to enhance physical component design. The updated Deep Think excels in rigorous academic challenges, achieving impressive scores on various benchmarks. It set a new record of 48.4% on Humanity’s Last Exam and reached 84.6% on ARC-AGI-2, verified by the ARC Prize Foundation. In competitive programming, it attained an Elo rating of 3455 on Codeforces and is expected to perform at a gold-medal level in future International Math Olympiads. Beyond mathematics, it shows strong results in chemistry and physics, including a 50.5% score on the CMT-Benchmark. Deep Think not only enhances academic research but also supports practical engineering applications. It can convert sketches into 3D-printable designs, analyzing complex shapes and generating the necessary files for production. Access to the updated features begins today for Google AI Ultra subscribers, and the early access program for the Gemini API is now open for interested researchers and enterprises.