Google said on the 19th (local time) that it released Gemini 3.1 Pro, which more than doubles the reasoning performance of its Gemini 3 Pro AI model.
Google said the new model was "designed for high-difficulty tasks that require advanced reasoning." It is a model for tasks where a simple answer is not enough, and it cited as strengths the ability to visualize and explain complex topics and to integrate and analyze diverse data.
In fact, the model scored 77.1% on ARC-AGI-2, a benchmark that measures AI's logical reasoning ability, higher than the 31.1% of Gemini 3 Pro. It also outscored OpenAI's GPT-5.2 (52.9%) and Anthropic Claude Opus 4.6 (68.8%).
On SWE-bench Verified, which measures coding ability, it scored 80.6%, rivaling Opus 4.6 (80.8%), which is rated at the top level. On "Humans' Last Exam" (HLE), which assesses academic reasoning, it also scored 44.4% without tools, ahead of GPT-5.2 (34.5%) and Opus 4.6 (40%).
Gemini 3.1 Pro will be offered first starting today to paid subscribers of AI Pro and AI Ultra.