Gemini 3 Deep Think has received a significant upgrade, positioning it as a specialized reasoning mode engineered to address complex challenges across science, research, and engineering domains. Developed in close collaboration with scientists and researchers, its core methodology is designed to elevate reasoning through profound mathematical and algorithmic rigor, particularly in scenarios characterized by a lack of clear problem definition, the absence of a single correct solution, and the prevalence of messy or incomplete data.

The system integrates deep scientific knowledge with practical engineering utility, moving beyond theoretical abstraction to enable tangible applications. While the underlying algorithmic architecture is not detailed, its operational methodology is described as one that processes complex information, applies sophisticated reasoning, and generates actionable outputs. For instance, it can interpret abstract input, such as a sketch, to model intricate three-dimensional shapes, subsequently generating a file for 3D printing, thereby transforming conceptual designs into physical realities. This implies capabilities in geometric reasoning, physical simulation, and generative design.

Key advancements and demonstrated capabilities include:

• Problem-Solving in Data-Scarce Fields: It successfully identified a subtle logical flaw in a highly technical mathematics paper within high-energy physics, a field with limited training data, a task that had previously eluded human peer review.
• Optimization and Design: It optimized fabrication methods for complex crystal growth, successfully designing a recipe for growing thin films exceeding $100 \text{ } \mu m$ in thickness to meet precise material targets. It also accelerates the design of physical components.

Its performance is quantitatively benchmarked across rigorous academic and competitive programming challenges:

• Achieved a new state-of-the-art result of $48.4\%$ on Humanity’s Last Exam (without tools), a benchmark designed to test the limits of frontier models.
• Attained an unprecedented $84.6\%$ on ARC-AGI-2, as verified by the ARC Prize Foundation.
• Reached an Elo rating of $3455$ on Codeforces, a competitive programming platform.
• Demonstrated gold-medal level performance on the 2025 International Math Olympiad.
• Achieved gold-medal level results on the written sections of both the 2025 International Physics Olympiad and the 2025 International Chemistry Olympiad.
• Showcased proficiency in advanced theoretical physics with a score of $50.5\%$ on the CMT-Benchmark.

Gemini 3 Deep Think is now accessible to Google AI Ultra subscribers through the Gemini app. Furthermore, it is being made available via the Gemini API to select researchers, engineers, and enterprises through an early access program.