Google just redefined the state of the art with a model that effectively doubles the reasoning performance of its predecessor.
Gemini 3.1 Pro has officially launched, and the capabilities are frankly startling. This industry pro took a deep dive into the release, revealing that the “Deep Think” model everyone has been raving about was actually running on this 3.1 engine the whole time. The standout feature isn’t just raw text generation; it is the model’s ability to understand complex spatial and physical concepts through code.
The expert highlights that this release isn’t a minor update; it’s a massive leap in reasoning capabilities designed for tasks where a simple answer fails.
📌 The “Pelican Bench” Proves Visual Coding Mastery
The creator of the video showcased something called the “Pelican Bench.” It sounds funny, but it is a rigorous test of a model’s ability to write SVG code that draws a pelican riding a bicycle. While previous versions created static or nonsensical images, Gemini 3.1 Pro generated a fluid, animated SVG of the bird in motion.
This matters because the AI isn’t just generating pixels; it is writing code that simulates physics and movement. The expert showed further examples, including a giraffe driving a tiny car and a complex “murmuring” simulation of bird flight patterns. This proves the model has a deep grasp of visual reasoning translated into executable code.
📈 Smashing the “Impossible” Benchmarks
The most shocking data point shared by the author involves the ARC AGI benchmark. This specific test measures a model’s ability to learn new skills and generalize information quickly, rather than just regurgitating memorized data. Gemini 3.1 Pro scored a massive 77.1%.
To put that in perspective, that is more than double the score of Gemini 3 Pro. The expert noted that it also outperformed heavy hitters like Opus 4.6 and GPT 5.2 in these tests. It also dominated on “Humanity’s Last Exam” and scientific knowledge tests, cementing its status as a powerhouse for logic and learning.
🏗️ From Text to Physical Reality (CAD & Urban Planning)
The practical applications demonstrated by the innovator go far beyond chatbots. He highlighted a use case from Google’s Jeff Dean involving an urban planning application that generates entire city layouts based on geographic constraints. You can adjust water features and paths, and the AI rebuilds the city block logic instantly.
Even more exciting for makers is the “Prompt to CAD” capability. The expert showed how you can feed the model a technical spec, and it generates manifold CAD code ready for 3D printing. This effectively bridges the gap between digital prompting and creating physical objects.
If you want to see the animated pelican or the city builder in action, you really should check out the full breakdown!