Google’s new Gemini 3.1 Pro AI model smashes benchmark records

Google has officially launched its latest large language model, Gemini 3.1 Pro, marking a significant advance in core reasoning and multimodal intelligence, the company and industry analysts say. The update more than doubles reasoning performance compared with its predecessor.

According to Google, Gemini 3.1 Pro achieved a verified score of 77.1 per cent on the ARC-AGI-2 benchmark, a demanding test designed to evaluate a model’s ability to solve entirely new logical problems it has never seen before. The result — more than twice the performance of the earlier Gemini 3 Pro — highlights improvements in abstract reasoning and complex problem-solving.

Gemini 3.1 Pro builds on the architecture of the Gemini 3 series, which Google first introduced in late 2025 as part of its strategy to push AI into deeper reasoning and more advanced real-world applications. The model’s enhanced capabilities extend across text, image and multimodal inputs, with a broad context window that allows it to handle long documents and integrated datasets.

The upgraded model is being made available in preview to developers and enterprises through the Gemini app, NotebookLM, Vertex AI, the Gemini API and other Google Cloud tools. Google says the expanded rollout aims to bring the elevated reasoning and synthesis capabilities of 3.1 Pro into more consumer and business workflows before its general release.

Industry observers note that Gemini 3.1 Pro’s benchmark performance places it ahead of many leading commercial AI models, including versions of OpenAI and Anthropic competitors on select reasoning tests, underscoring the competitive dynamics in advanced AI research.