Google has launched Gemini 3.1 Pro, a new artificial intelligence model demonstrating superior performance in advanced reasoning tasks, surpassing rivals like Claude 4.6 Opus and GPT-5.2 on key benchmarks.
This Transformer model utilizes a "mixture of experts" architecture, optimizing parameter activation for prompt responses. Gemini 3.1 Pro supports prompts up to one million tokens, incorporating multimodal data including video, and generates responses up to 64,000 tokens.
On the challenging ARC-AGI-2 benchmark, which assesses an AI's ability to deduce patterns from visual puzzles, Gemini 3.1 Pro achieved a score of 77.1%. This performance places it approximately 24% ahead of GPT-5.2 and nearly 9% ahead of Claude Opus 4.6, following tests conducted in a hardware-intensive mode designed to enhance reasoning capabilities.
Google reports that Gemini 3.1 Pro has also set new records on other benchmarks, including MCP Atlas for third-party service integration and Terminal-Bench 2.0 for coding. It further outperformed Claude Opus 4.6 on the scientific programming benchmark, SciCode.
Demonstrations showcase Gemini 3.1 Pro's versatility, including generating an HTML visualization of Earth's orbit using real-time ISS location data and creating a website from a novel. The model also exhibits significant improvements in generating Scalable Vector Graphics (SVG) files, a format crucial for web applications due to its scalability and interactive potential.
Gemini 3.1 Pro is accessible as a preview in Google's development tools and to consumers via the Gemini app and NotebookLM. Enterprise users can leverage its capabilities through Google's Vertex AI suite.
The "upgraded core intelligence" powering Gemini 3.1 Pro debuted last week in Gemini 3 Deep Think, a processing mode designed for scientific tasks. This enhanced feature has already aided early users in identifying a flaw in a mathematics paper and in the manufacturing of new semiconductor structures.