Google Gemini 3.0 Pro Is CRAZY! It Turned My Sketch Into a Real Game…

It finally happened. After weeks of leaks, screenshots, and tweets on X, and everyone refreshing their Google AI Studio accounts, Gemini 3 is officially live. We've been tracking every hint from the early model IDs buried into the code to the internal docs that accidentally slipped out to the devs posting vague eyeball emojis. And now that it's actually here in AI Studio, we can finally see what Google's been cooking behind the scenes. Gemini 3 isn't just an upgrade. This is the version people have been waiting for. The one that's supposed to push Google back into the conversation with real reasoning improvements, faster responses, stronger coding abilities, and the multimodal performance they've been hyping for over a month now. Today, I'm jumping straight into AI Studio to test it live, go through what's new, what feels different, and whether it actually lives up to the leaks. So, let's get into it. All right. Before we

get into the demo, let's start with the official Gemini 3 Pro page because Google came out swinging here. Right up top, they're pitching this as the best model for Agentic workflow and vibe coding. Basically saying this is their most creative, most autonomous version yet. And I love that phrasing, bring creative concepts to life. This is Google saying give Gemini 3 an idea and is going to run with it. And right underneath that, they're actually highlighting something we all have been waiting for. actual reasoning power. Not just sounding smart, not giving you fluffy answers, but real planning and real multi-step breakdowns and real execution. Then, as you scroll down, they hit you with this line, partner with a pro. This is a bold shift in tone. This is Google positioning Gemini 3 Pro as something closer to a collaborator than a chatbot. And it breaks the capabilities into three pillars. Learn anything so you understand complex topics with clear, concise explanations. Build anything. Turn your sketches, ideas, or half-written prompts into real tools. And lastly, plan anything. Delegate multi-step tasks. The backbone of Agentic AI. This is their way of saying Gemini 3 isn't just reactive. It knows how to run with something. Then when you hit the hands-on section, this is where Gemini 3 really separates itself. The first example is pure vibe coding. The model generating beautiful responsive visual experiences. This is the front-end code that actually looks good, not the bootstrap looking stuff we were getting last year. Then the next preview is all about deeper understanding. And this is where Google is flexing the reasoning upgrades, showing simulations, interactive content, and nuance breakdowns. Gemini 3 is trying to be the model that doesn't just answer your question, but actually teaches you. And the last one, it's super interesting. It shows that Gemini 3 generating interactive flash cards and even games to help you learn faster. So this is where multimodal comes in. It pulls from text, images, video, audio, and even code at the same time to build custom mini apps or solutions that we see in this example. Then you get to the part of the page that everyone has been talking about, the benchmark charts. And these numbers are insane. So let's start with the humanity's last exam. Gemini 3 Pro scores 37. 5% while the new GPT 5. 1 is at 26. 5% which is a huge reasoning gap. Then we have the terminal bench 2. 0 which is the agentic coding capabilities. Gemini 3 pro is at 54. 2 and GPT 5. 1 47. 6. This is literally Google showing yes our model can run terminal task fix code and act like an actual dev assistant. Then we have the simple QA verified which is testing the parametric knowledge. Gemini 3 Pro is at 72. 1% while Gemini 2. 5 their older model is at 54. 5% and GPT 5. 1 sits at 34. 9%. So this is a massive jump in factual accuracy. Now if you were interested in looking at the full benchmark table you they have also included in here and you can look at Gemini 3 leading in reasoning math multimodal agentic coding chart understanding screen understanding OCR the one big headline screen spot pro Gemini 3 hits 72. 7% and GPT 5. 1 only gets 3. 5% that's basically domination by Gemini the bottom they summarize everything into four tiles number one reasoning with unprecedented depth and nuances which basically means smart concise direct responses without becoming fluffy. Then worldleading in multimodal understanding so text imagines video audio even code. Gemini 3 is the state-of-the-art model for this. Then they're claiming it to be the best model for vibe coding and agentic coding which basically means Gemini 3 brings exceptional instruction following and it also allows you to improve tool use and agentic coding. And lastly, we have Agentic capabilities, which means you have the ability to use Gemini 3 for better tool use, multi-step tasks, and build intelligent personal AI assistance that not any other model could do. All right. So, once you're in

Google AI Studio, you can try the new model here. So, I'm going to click on try Gemini 3. So, Google keeps promoting Gemini 3 Pro as the best model for vibe coding. meaning you can take random creative ideas, sketches, doodles, whatever, and it's supposed to turn them into an actual working app or a game. So, I made this extremely professional sketch, as you can see here, and I'm trying to make it create a game where there's a guy, he's running through obstacles and dodging stuff at the top. And we're going to see if Gemini 3 Pro can actually turn this into a real game. I have given it the image. I'm going to tell it this prompt, which is basically, here's a sketch of a game idea I had. Turn this into a fully working interactive browser game using HTML, CSS, and JavaScript. Please include animation, sound effects if possible, a scoring system, a win and loss condition, clean code in separate files, and make the visuals fun, modern, and playful. Also, generate instructions so I can run this locally. So, this is the picture, and that's the prompt. Let's see what it does. It's analyzing the scene. So, let's just see what it says for that. I've been dissecting the sketch, focusing on the stick figure, the platforms, and the ball's trajectory. Currently, I'm interpreting the context and implied actions within the scene, including the date mentioned on the image. I just put the November 18th date today just to show you guys that this is a sketch I created right now, but I'm trying to identify potential game mechanics or challenges based on interplay of these elements. Okay, so it's calling it the infinite runner. Okay, that's genre has decided based on this based on the image with the player controlling the stick figure. Not bad. Thinking dodging flying balls. So, that's pretty good. It kind of understood my amazing sketch and understood what it's supposed to do. Uh, it's designing the obstacles, outlining the game structure, defining visuals and logic, defining obstacle details. So, okay. So, it's doing a lot of thinking and it's going to create an infinite runner game. And it's I've interpreted your sketch into a game called Doodle Dash. So, it's calling the game Doodle Dash. You play as the stick figure. You must jump over the benches and duck under the flying balls. Okay, so it has actually done generating this already, I think. Wait, has it? Okay. Yeah, it's done. And it has show me how I can run this locally, which is create a folder on your desktop called Doodle Dash. And then inside the folder, create three text files named exactly this, this. Copy the code provided above into the respective files. Play. And if I want to play this, I have to click the doubleclick index. html to open it in your web browser. And it has given me instructions on how to play this. What's crazy is this only took Gemini 20. 4 seconds. Now, obviously, we haven't seen what it looks like. So, let's get into that. Okay. So, I didn't want to download the files and do all that work. So, I just asked it to create here. And it said that yeah, I can do this everything. It's going to merge the HTML, CSS, and JavaScript into one block. Um, it's adding the date. Okay. Not bad. And I can click the start game below to play it. So, where is the start game option? Okay, it's still generating it. Okay, guys, this is amazing. Look at the design of it. It put the November 18 date. It also has the start game button, which looks amazing. Space or plus to jump over benches and the below arrow. Let's start the game. Oh, I'm really bad at it. Let's do it again. Got it. Okay, this is actually pretty fun. I ain't going to lie. Okay, this is actually not bad for what I was able to create. It has the score system going. It has it literally took drawing that I gave it. Like if you look at the ball, it looks like my original image as well. So, it's not bad at all. And this was done in like 20 seconds. And I'm able to deploy it in here. So, let's die. It keeps my score count and I can do try again and run it. So, this is really cool, guys. And this is only with the model being released right now. This model is supposed to get better and better. So, let's see what the future looks like. But if this is what we're at right now with Gemini 3, I can't imagine what the upgraded model is going to look

like. If you enjoy this video, this is what we do here. Fast, clear updates on the biggest moves in AI. If you want to stay ahead of everything happening in this space, make sure you're subscribed. And if you want the hands-on side, demos, tools, workflows, and everything developers can actually build, well, check out the World of AI. We also run a simple, no noise newsletter that gives you the most important AI tools and updates in just a couple of minutes. Subscribe here. Follow World of AI. Join the newsletter.