Google Gemini AI recreated working Windows 11 code from scratch

On February 20, Google officially introduced Gemini 3.1 Pro. According to the company, the new version outperformed Gemini 3 Pro, Claude Opus 4.6, Claude Sonnet 4.6, and GPT-5.2 in a number of internal tests.

One of the demonstration cases was a WebOS copy of Windows 11, completely generated by the model in the browser. The interface includes program icons, the Start menu, and the basic logic of window operation. This is not a full-fledged operating system, but an interactive simulation with a simplified structure.

Blogger Chetaslua showed how the model assembles such an environment within the framework of an agent system. He previously tested a similar project on Gemini 3.0 Pro, but the new version generates a more complex interface and adds elements of interaction between components.

Test results and limitations

According to Google, Gemini 3.1 Pro took first place in 12 benchmarks that tested logical thinking, multimodality, and working with a long context.

At the same time, in SWE-Bench Pro and SWE-Bench Verified, tests focused on real engineering tasks, the indicators were lower. These scenarios involve working with large code bases and fixing bugs in existing projects, which requires deeper integration and resilience to complex code structures.

The main update of the model concerns complex reasoning and performing engineering tasks. Google positions Gemini 3.1 Pro as a tool for complex technical and creative projects.

Interactive projects: from SimCity to 3D simulators

In demos, the model also creates interactive city-building simulator-level designs: it generates an infrastructure map, simulates traffic, and visualizes the urban environment.

Other examples include a 3D flight simulator of a flock of birds with gesture control and a generative soundtrack, as well as an interactive model of the Solar System with illuminated planets and the ability to move between them.

Access to the model

Gemini 3.1 Pro is available to Google AI Pro and Ultra subscribers in the Gemini and NotebookLM apps. Free users are given two requests for the new model.

Developers can access it through AI Studio, Vertex AI, and the Google API.