Gemini 3 Pro Unleashed: From Instant Webpage Replication to Record‑Breaking AI Benchmarks

Google’s Gemini 3 Pro has finally arrived, and the author spent a late‑night session testing its capabilities, reporting that the model feels dramatically more powerful than its predecessor.

Webpage replication : Using a simple prompt—"Please help me replicate this website homepage"—the model reproduced the homepages of Bilibili, X (formerly Twitter), Douyin, and Xiaohongshu. The results are shown in the following images:

Visual generation improvements : The author compares weather‑card images produced by Gemini 3 Pro and Gemini 2.5 Pro, showing clearer graphics and richer detail in the newer model.

The model also excels at dynamic visual tasks: a six‑sided bouncing‑ball test runs smoothly, and a single prompt can generate a fully playable Plants‑vs‑Zombies‑style game.

Another prompt demonstrates 3‑D animation generation:

Prompt: Use HTML to generate a 3D demonstration animation of DNA double‑helix replication, like a high‑school science presentation.

Remarkably, the author got Gemini 3 Pro to write a complete macOS‑style operating system with a built‑in Gemini AI assistant that can converse and respond intelligently.

Uploading an architecture diagram allows the model to recreate the page and edit the embedded text directly, making diagram updates much easier.

Beyond visual tasks, Gemini 3 Pro can generate functional applications. The author asked it to create a public‑account cover‑image generator that takes article content as input. The exact prompt is shown below:

帮我做一个公众号封面图生成器，只需要输入公众号文章内容，自动根据内容帮我生成符合要求的公众号封面图  
你可以参考图片中的样式进行设计开发

Performance metrics highlight the model’s strength: it achieved a score of 1501 on LMArena, topping the leaderboard; it scored 37.5% on the “Humanity’s Last Exam” test; and it set new records on MathArena. The new “Deep Think” mode targets complex reasoning problems.

Google also released an Agent development IDE called Antigravity , offering free access to Gemini 3 Pro and Claude Sonnet 4.5.

On code‑generation benchmarks, Gemini 3 Pro reached 76.2% on SWE‑bench, indicating strong potential for project‑scale development. Its context window remains at 1 million tokens, and it can analyze a 20‑minute video frame‑by‑frame.

The model is already accessible via Google AI Studio, the Gemini app, Gemini CLI, and the Cursor platform, and its API can be integrated into custom business systems.

While Gemini 3 Pro represents a leap forward for front‑end engineers—potentially replacing those who only write code—it still cannot fully replace senior engineers who design architecture and understand business logic. Many see this release as a milestone toward AGI, echoing the excitement that followed GPT‑4’s launch.