A Quantum Leap in AI Logic
Google has officially launched Gemini 3 Deep Think, a sophisticated new reasoning mode now available exclusively to Google AI Ultra subscribers within the Gemini app. Representing a significant architectural evolution, this update is engineered to solve high-friction problems in mathematics, science, and advanced logic that have historically stumped even the most capable large language models (LLMs). By moving beyond simple pattern matching, Deep Think introduces a more deliberate, thoughtful approach to artificial intelligence.
Shattering Industry Benchmarks
The performance data accompanying this release signals a major shift in the competitive landscape. Gemini 3 Deep Think has established new industry standards on some of the world’s toughest AI evaluations. It achieved a striking 41.0% on Humanity’s Last Exam without external tools—a test specifically designed to be difficult for AI. Furthermore, it reached an unprecedented 45.1% on the ARC-AGI-2 benchmark (utilizing code execution), a rigorous assessment of general artificial intelligence that measures an agent's ability to adapt to novel reasoning tasks.
The Power of Parallel Reasoning
The secret behind this leap in capability lies in advanced parallel reasoning. Unlike traditional linear processing, Gemini 3 Deep Think explores multiple hypotheses simultaneously, weighing various intellectual paths before converging on the most accurate solution. This methodology is a direct evolution of the experimental Gemini 2.5 Deep Think variants, which recently proved their mettle by securing gold-medal standards at prestigious human competitions, including the International Mathematical Olympiad and the ICPC World Finals.
How to Access Gemini 3 Deep Think
Starting today, the new mode is fully integrated into the Gemini ecosystem for premium users. Google AI Ultra subscribers can activate the feature immediately by navigating to the Gemini app, selecting "Deep Think" from the prompt bar, and ensuring Gemini 3 Pro is chosen in the model dropdown menu. This seamless integration places the power of world-class computational reasoning directly into the hands of developers, researchers, and power users.
