Google today announced a major upgrade to Gemini 3 Deep Think. This is an inference model designed to solve scientific, research and engineering challenges. Currently, Google AI Ultra subscribers can already use the updated version in the Gemini App. At the same time, Google has opened Deep Think early access to the Gemini API to selected researchers, engineers, and enterprises for the first time.
From abstract theory to hard-core engineering
The core of this update lies in the evolution of "deep thinking" capabilities. Gemini 3 Deep Think not only strengthens the scientific knowledge base, but also focuses on solving complex problems that lack clear boundaries and have confusing or incomplete data.
The Google DeepMind team says the model aims to bridge the gap between esoteric scientific theory and everyday engineering applications.
In the specific performance evaluation, Gemini 3 Deep Think showed outstanding data.
On the Humanity’s Last Exam benchmark, which tests the limits of modern cutting-edge models, the model achieved 48.4% accuracy without the use of additional tools, setting a new industry standard.
Even more interesting is its performance in the ARC-AGI-2 test. Verified by the ARC Prize Foundation, Gemini 3 Deep Think achieved an unprecedented high score of 84.6%.
This achievement is of great significance in the field of general artificial intelligence reasoning. Additionally, the model achieved an Elo rating of 3455 in the highly competitive Codeforces programming challenge and reached gold medal level at the 2025 International Mathematical Olympiad.
Practical cases that go deep into the front lines of scientific research
Unlike previous models that only focused on running scores, Gemini 3 Deep Think has entered top university laboratories.
Mathematician Lisa Carbone of Rutgers University applied the model to a review of mathematical structures in high-energy physics.
In a field with very little training data, Deep Think successfully reviewed a highly technical mathematics paper and identified a subtle logic flaw that had previously gone undetected through human peer review. This discovery has potential value in connecting Einstein's theory of gravity with quantum mechanics.
In materials science, Duke University’s Wang Lab uses Deep Think to optimize manufacturing methods for growing complex crystals. The model successfully devised a recipe for growing films larger than 100 micrometers (μm). This is an extremely precise goal, and previous methods often fail to respond to this type of challenge.
The ability to build the physical world
In addition to theoretical calculations, Google also demonstrated the model's potential for physical manufacturing. Anupam Pathak, director of research and development for Google's Platforms and Devices group, tested the new version of the model's ability to accelerate the design of physical components.
The demonstration showed that Deep Think has the ability to convert hand-drawn sketches into physical objects.
Users simply provide a sketch and the model analyzes the drawing, models complex 3D shapes, and generates files ready for 3D printing.
This means that the transformation process from concept to physical prototype is greatly compressed.
Comprehensive coverage of scientific fields
Beyond math and programming, the Gemini 3 Deep Think excels in a wide range of scientific fields, including chemistry and physics. According to data published by Google, the model achieved gold medal levels in the written portion of the 2025 International Physics Olympiad and Chemistry Olympiad.
In the CMT-Benchmark test for advanced theoretical physics, it achieved a score of 50.5%, showing proficiency in handling complex scientific areas.
Currently, Google AI Ultra subscribers can experience this upgrade in the app starting today. For developers and businesses who want to integrate this capability through APIs, Google has opened an early access application channel. This update marks the accelerated transformation of AI models from general chat assistants to professional scientific research and engineering assistance tools.