Gemini 3 Pro GA: Google's 4 SECRET Gemini 3 Pro Models LEAKED!

Anthropic dropped Cloud Opus 4. 6 this week. OpenAI pushed their updates with GPT 5. 3 Codeex and Google's been secretly testing Gemini 3 Pro G8 in the arena. And I just found out that there are actually four leaked variants of Gemini 3 Pro G8 being tested right now. It's not announced and it's not officially released. Just quietly testing in arena battles while everyone thinks Google is behind. I've got the leaked generations from these models and we're going to break down exactly what Google is doing. So, let's get into it.

The four Gemini 3 Pro GA variants currently in testing are Riftrunner, which is the standard Gemini 3 Pro, Snowplow and Snowball, which are both Gemini 3 Pro, but two separate instances, Fire Falcon, which is the Gemini 3 Flash variant, and Nova Quida, which is another Gemini 3 Flash variant. There are two checkpoints that are heavily biased towards SVG generation and apparently worse at everything else. And here's the interesting part. One checkpoint was better at everything and Google pulled it away from testing. Why test two instances of the same model? Why bias checkpoints towards SVG and why remove your best performing variant from the testing? Multiple variant means Google is AB testing different approaches before the GA release. It's smart, but also tells us that they don't know which configuration is best yet. The two snowplow and snowball instances could be different training methods, different approaches, or different safety configurations. They're testing the same base model with variations to see what performs better in real usage. SVG bias checkpoints. This happens when you optimize too hard for one capability. The model gets better at SVG generation but worse at general task. Google is trying to find the balance. The remove checkpoint is the real story here though. They had a variant that performed better across the board and pulled it. Possible reasons, safety concerns, alignment issues, maybe too expensive to run at scale or maybe failed internal red testing. Also, the Flash models are untested yet, which means Google isn't ready to show their efficiency play yet. Flash is supposed to be faster and cheaper, and if they're holding those back, they might not be confident in the spend to quality ratio yet. Now, I want to show you guys actual

generations from these Gemini 3 Pro variants. The first one you're seeing is a SVG generation. I know it doesn't look amazing, but let me explain what SVG is just in case you don't know. SVG is scalable vector graphics, which is an XMLbased open standard vector image format used for 2D graphics that remain crisp at any resolution. Unlike other formats like JPEG, SVGs actually use mathematical formulas for lines and shape. So, this is all programmed. Everything that you're seeing is programmed by math and the image generations are going to stay consistent depending on whatever size it is. So if it's a simple logo, it'll look the same, but if you blow it up on the screen, it will look the same unlike JPEGs which would change in resolutions. So what you're seeing right now is the SVG generation of the Mona Lisa from the Gemini 3 Pro variant. Obviously when you look at it, this is not close to the Mona Lisa, but like it's pretty good. And the reason why I'm saying it's pretty good and pretty better than the other models is because I'm going to show you right now what clot opus 4. 5 and GPT 5. 3 created. We're clearly seeing some humanlike qualities in the eyes and everything like that. The nose is getting there. Obviously, these models are still in the early stages, but now let me show you what the other generations look like. So, these are the generations from GPT 5. 3 Codeex, the new one, and Opus 4. 6, which is the newest model right now. You can see. Okay, this looks like not Mona Lisa. I feel like this person would get cancelled if it was in Karn's culture because what is going on here? I have no clue. Right. And then we look at the one by Opus 4. 6. It looks like something a And why is her clothes ripped? Like what is the model doing over here and it looks like a nun rather than the Mona Lisa? But interesting, right? So clearly after you see these two, you are going to appreciate this SVG generation by Google because this one is obviously much better than the other two variants that we just saw. And it shows that obviously Google was trying to go hard on the SVG generation. I don't know why, but we can clearly see that it's actually translating into real use cases. And just to show you how good Google is trying to get at SVG generation, we can see this clearly in this example over here. So this image over here is Claude Opus 4. 6. So this is a famous test which is a pelican riding a bike. So all the new models are being tested on it. But it looks like Gemini 3 Pro G8 obviously has the best result over here. And this model is probably going to solve this test. Meaning that we won't be seeing this generation or this test being tested going forward because obviously it looks like Google Gemini 3 Pro G is getting there. Right. Cloud Opus 4. 6 the newest model. It's not even there yet. And then we see this version over here, which is their old version, which is actually not bad, but clearly this is the winner. Okay, what I'm going to show

you right now is something that this person from Twitter created using the new Gemini League models. Just a fair warning, it's pretty intense visually. It's kind of like a psychedelic, so maybe skip if you're sensitive to that type of stuff, but put on headphones and watch this trippy visual thing for about 2 minutes. It's got flashing geometric patterns and binaural beats that explore what would AI consciousness might feel like. So, let's get into it. —

— So, here's another generation which is a complex 3. js 3D graphics prompt. The prompt here is pretty technical, but is basically asking the model to create a glass dot header drawn, which is a 12-sided 3D shape where each face acts like a filter or a mask. In the center, there's a Taurus knot rotating. And when you look through the different faces of the model, you see the same knot, but it looks different through each face. Then there's a second larger glass shell around the whole thing. And each face of this outer shell changes how the inner shape appears. One face makes it look like a tinted glass. Another like a stained glass, another shows it as a wireframe. And the technical requirements are pretty specific. Single HTML file, 3. js library, geometry based rendering set to zero, so you don't get duplicate reflections, opacity between 0. 1 and 0. 3. Obviously, I didn't make this prompt. This is actually from Subwoofer. I guess he really loves geometry. But this kind of shows you how good this model is. So, if we take a look at the shape, like it looks pretty accurate. We see a 12-sided shape. And then if we go here, we see a different version versus here versus here. So the prompt is actually followed. And then if you look at the inner knot, which kind of looks cool like I don't even know what you call this shape. And I don't even know how you think of this prompt. So subwoofer must be some sort of genius because this is actually insane. And it's able to hold this type of consistency, right? Like if I go here, I'm going to see the same thing if I go here. So, the shape is pretty accurate and everything looks really coherent based on the prompt given. But I can't wait for this new model to come out. Well, hopefully Google doesn't nerf this because most of the times when Google is testing these models on Arena, we've seen amazing results. But as soon as it comes out, for some reason, it's not the same because they usually end up nerfing the models before they release. So, let's pray that Google doesn't do that this time around cuz if we get access to this, Google might take number one on the leaderboard once again. Make sure to

subscribe to our channel. We do real tests, not just headlines. Make sure you're also subscribed to the world of AI. And don't forget to check out our newsletter for deeper breakdowns you won't see on YouTube. And I'm growing my Twitter following, so make sure you follow me on Twitter as well. — Hope you guys enjoyed today's video and I'll see you in the next