Google AI's biggest bombshell this year. The test exposure shows it can directly replicate macOS. It's more worth looking forward to than GPT-5.

Remember the wave of Google AI suite we introduced not long ago? At that time, nano banana was extremely popular, and Gemini topped the list. Many people tried Gemini 2.5 Pro for the first time and found that this low - key flagship model was actually better than GPT - 5 in many cases.

Now, the more powerful Gemini 3.0 has been “prematurely launched” in the Google AI Studio test.

Netizens have found that its ability has improved significantly. It can directly generate a web - based Windows, macOS, etc. with a simple set of prompts.

A large number of test results mention that the code written by Gemini 3 can run directly, and the front - end design is more mature. It can even create tools for video editing and short - video browsing.

This could be the moment when Gemini surpasses Claude and GPT - 5 again, beyond the image - editing field.

However, whether it's just marketing like GPT - 5 at that time or can really change our user experience, let's first take a look at these cases of Gemini 3 shared by netizens.

Starting with writing a Windows?

Just as nano banana uses the performance of a model in text generation as an indicator when judging image - generation ability.

And when it comes to judging the delivery ability of a foundational model, programming, a project where even a slight error in the code can make it unrunnable, has become the focus of most foundational model manufacturers. Claude, which performed well in this area at the beginning, also gained some popularity because of this.

After selecting Agentic Coding Average, in the model ranking list, Claude Sonnet 4.5 ranks first. Source: https://livebench.ai/

Recently, several bloggers on X have reported that the upcoming Gemini 3's real performance in multiple programming tasks is not only expected to beat GPT - 5 but also shows a trend of directly surpassing Claude.

Logan, the person - in - charge of Google AI Studio, also replied to a netizen on X, confirming that Gemini 3 has the ability to call tools, which is what we commonly call an agent.

Previously, most programming tests focused on understanding logical ability, such as a physics test of a ball flowing inside a hexagonal frame; aesthetic judgment of front - end ability, observing whether the website generated by AI is user - friendly and has a modern page; and some practical mini - games or tools.

This partial release of Gemini 3 tests can be said to be a major upgrade of the test samples.

X blogger chetaslua shared multiple programming tasks completed by Gemini 3. It can directly generate a web - based macOS computer for us. In this web page, it's like operating a cloud computer. All applications can be opened. Clicking on the Safari icon can open the browser and continue to browse the web.

prompt: Design and create a web os like macOS full functional features from text editor, terminal with python and code editor and a game that can be played to dile manager to paint to video editor and all important windows os pre bundled software Use whatever libraries to get this done but make sure I can paste it all into a single HTML file and open it in Chrome.make it interesting and highly detail, shows details that no one expected go full creative and full beauty in one code block Prompt: Design and create a fully - functional web - based operating system similar to macOS, starting from a text editor, a terminal with Python, and a code editor, and including a playable game, file manager, painting tool, video editor, and all important pre - installed software of the Windows operating system. Use any libraries to complete this task, but ensure that I can paste all the content into a single HTML file and open it in Chrome. Make it interesting and highly detailed, showing details that no one expects, and achieve complete creativity and beauty in one code block

Source code: https://codepen.io/ChetasLua/pen/yyezLjN

Source: https://x.com/chetaslua/status/1977866953705316571

Not only can it generate a macOS cloud computer, but the blogger also modified the same prompt to generate operating systems such as Windows, Linux, and cloud phones. The boot animation and in - system operations can also be perfectly restored.

Moreover, the speed of Gemini 3 in handling these programming tasks has also been improved. According to the blogger's reply in the comment section, the inference speed of Gemini 3 is 1 - 2 minutes.

When the same prompt was given to Claude Sonnet 4.5, netizens shared that Claude only generated a page, and all applications couldn't be opened. I also tested the ability of GPT - 5. Although it could open some applications, there was no further processing logic after opening them.

https://x.com/pallavmac/status/1977979879489679640

There's also a little Easter egg. In the cloud computer opened on the web page, click on the terminal, enter “matrix”, and we'll see a flowing 01 code similar to the poster of The Matrix movie. It can really be called the “Matrix Revolution”.

Besides these popular system - restoration tests, the programming ability of Gemini 3 is also reflected in its front - end, which has made progress that other AI models can't achieve. The most obvious one is that Gemini 3 has finally got rid of the gradient purple that often appears in AI programming.

x@Liam06972452, prompt: Write full HTML, CSS and JavaScript for a stunning, bold, creative and unique landing page for a website about llms. It should look like it was built from scratch by a world - leading web design agency.

Just like for this website about large language models, the input prompt was simply “Write complete HTML, CSS, and JavaScript for a website about LLMs to create a stunning, bold, creative, and unique landing page. It should look like it was built from scratch by a world - leading web design agency.”

Netizens have said in the comment section that Gemini 3 must be collaborating with designers. Some people predict that Gemini 3 will dominate the front - end field, just like nano banana.

However, some people think there are still some problems with the front - end of Gemini 3. Although it has solved those annoying gradients and changed what most AI models have been doing for a long time, the front - end of 3.0 is always black unless specifically prompted. In my opinion, black is much better than gradients, and website design should be simple and elegant.

Based on its ability to create an operating system with a single sentence and its excellent front - end ability, Gemini 3 can also generate an online video editor, which can really edit videos.

Prompt: create amazing single html page video editor that WORKS! make it amazing. (Create an amazing single - HTML - page video editor that must work! Make it really great.) Source: https://x.com/sbalhatlani/status/1977501190322549017 CodePen runnable experience link: https://codepen.io/DiNaSoR/pen/WbrEELE

A web - based version of the Douyin app that allows infinite video browsing.

Source: https://www.reddit.com/r/Bard/comments/1o67jz4/crazy_google_gemini_30_pro_just_built_a_perfect/

Experience address: https://jsbin.com/yisixokuwi/1

Those tools like Lovable and Cursor, which are used for vibe coding and one - click generation of web applications, may need to consider changing the default model to Gemini 3.

In the classic code - ability test, there is also a task of generating a pelican riding a bicycle to compare the SVG code generation effects of different models.

Gemini 3 can generate an ordinary version of the pelican.

https://x.com/whylifeis4/status/1974205929110311134

There is even a 3D version in the same style as the game “Minecraft”.