Google LLC today introduced Gemini 3.1 Pro, a new reasoning model that outperforms Claude 4.6 Opus and GPT-5.2 across several benchmarks.

The algorithm is available via more than a half-dozen of the search giant’s products.

Gemini 3.1 Pro is a Transformer model with a mixture of experts architecture, which means that it activates only some of its parameters when generating a prompt response. Users can enter prompts with up to 1 million tokens’ worth of data, including not only text but also multimodal files such as videos. Gemini 3.1 Pro’s responses contain up to 64,000 tokens.

Google evaluated the model’s reasoning capabilities using ARC-AGI-2, one of the most difficult artificial intelligence benchmarks on the market. It comprises visual puzzles that each include a series of shapes. The shapes that make up a puzzle differ from one another in their design, but all follow a certain pattern. LLMs must deduce the pattern and use it to generate a new shape.

Gemini 3.1 Pro achieved an ARC-AGI-2 score of 77.1%, which put it about 24% ahead of GPT-5.2. It also outperformed Anthropic PBC’s Claude Opus 4.6 by nearly 9%. All three models were tested in a hardware-intensive mode that improves their ability to tackle reasoning tasks.

According to Google, Gemini 3.1 Pro also set records on several other benchmarks. The list includes MCP Atlas, which evaluates AI models’ ability to perform tasks using third-party services, and the Terminal-Bench 2.0 coding test. The model performed 7% better than Claude Opus 4.6 on another coding benchmark, SciCode, that comprises scientific programming tasks.

A demo published by Google shows Gemini 3.1 Pro generating an HTML visualization of the Earth’s orbit. The dashboard uses data from a third-party service to show the current location of the International Space Station. In another demo, the model created a website based on a novel.

Compared with the previous-generation Gemini algorithm, Gemini 3.1 Pro is significantly better at generating SVG files. SVG is an image format widely used in web application projects. SVG images often include interactive elements and can be resized it without resolution loss.

Gemini 3.1 Pro is available as a preview in several of Google’s development tools. Consumers, meanwhile, can access the model via the Gemini app and NotebookLM. Google is also bringing Gemini 3.1 Pro to several other offerings including its Vertex AI suite of AI cloud services for enterprises.

According to the company, the “upgraded core intelligence” that underpins the model made its initial debut last week in Gemini 3 Deep Think. It’s a processing mode that enables Google’s LLMs to tackle scientific tasks. The upgraded version of the feature has helped early adopters identify a flaw in a peer-reviewed mathematics paper and manufacture new semiconductor structures.

Image: Google