All about Gemini 3.1 whether it beats Claude 4.6 and what to use it for!

Hey everyone, hope you're all having a wonderful day. Welcome back to the Matt Vid Pro channel. Gemini 3. 1 has dropped. Google has crawled into the servers and drug out their latest beast. Gemini 3, when it first released set the standard for how models should behave with instant codability, actually being able to get something right on the very first try. 3. 1 now takes this a step further and similar to clawed opus 4. 5 and 4. 6. Let's say this model really feels like it has extra breadth and can impress you a little bit. It's animating a full SVG

of a train through the mountains. It's not just generating a single picture. This comes from the legendary Chetislua. Always testing out the latest LMS. XAI also recently dropped Groank 4. 20. It uses four Grock agents at once to deliver a final result. And tested on the same prompt. Honestly, I'm thinking that this looks quite a bit worse than the Gemini 3. 1 Pro. Yeah, it's got all four seasons. The tunnel is there, but it's just way more rudimentary. Like the train that Gemini 3. 1 Pro made went through the tunnel and had actual lighting on the barrel. Let's go ahead and try this out in a standard Gemini chat. All right, so it thought for maybe two and a half minutes. Fairly quick result. I can tell you for sure chat GPT Pro would take way longer, but Claude Opus 4. 6 is really where you'll see this model kind of trade blows back and forth. For a first pass result right through the web app, this is really, really impressive to generate full animated train SVG. I don't think this one looks as nice as Chedda's example, but I think I know why. This still looks better than the Gro 4. 20, but something

that I've noticed is that the Gemini web app interface really seems to be set up and designed for faster chats. I mean, it thought for less than 3 minutes to produce this. That's not a long-term agentic flow. And this LM can absolutely be used in an long-term agentic flow. So you guys will see that later. But in the Google web chat interface, just a little bit more lazy, like it's in a lowinking mode. In terms of vibes or a basic sniff test though, four chats, this actually goes pretty well for me. Gemini 3 always felt kind of sterile to me, but if you sent it the right meme, it would kick something good. And in this example, the meme is a little bit dark. It's also kind of specific and technological. After thinking for just a short amount of time, it gives me a nice modern human feeling AI response, not too much going on, here's exactly how this deeply cursed, cross-temporal conversation would play out in the afterlife waiting room. That extra juice you get from these incredible modern-day LMs, you know, the translation guide, you could export it to Sheets or something. But of course, honestly, it's even impressive that it understands this meme right off

the bat. The visual understanding and reasoning in this model is incredible. In fact, from this blurry image, let's have it identify these two actors and the movie scene. It thought for a little bit, but sure enough, perfectly correct. William Defoe, Christian Bale, and the movie itself. It's a very famous movie scene and memes, so we expected to get this right, but it is pretty crunchy. Let me take you through even more insane community demos. Andrea's here compares Gemini 3 Pro versus 3. 1 Pro. This is only 30 seconds long, but just seeing it side by side here. The animations, the level of depth, and the understanding of where different parts of the scene need to be placed. It's kind of incredible how well it's able to spatially understand a lot of these. That chameleon demo, look at the back legs here. understand it's writing code to actually formulate each bend and each leg have the ones that are faded behind even with shading on certain parts. Gemini 3 doesn't even come close. His smile is falling off his face. There's no shading at all and he's got the eye on the other side of his head is poking through. Legit API shows off uploading this heavily degraded image. It's supposed to be Peter Griffin and Brian. Our image recognition skills as humans are so incredible. We'll pick that up on the spot if we've seen the show and the characters, but for an AI model picking this up from like the blob of hair on Peter's head and the vague shape, to me, that is just crazy image recognition skills. So many people have also been

saying that this model is state-of-the-art in 3D spatial reasoning. And yeah, the demos do not lie. I am severely impressed by what the community has shown off, but I haven't really been able to replicate it for myself. Zack Dive shows off this 3D dirt bike reconstruction, and it is incredibly impressive. You can see the single cylinder engine. wheels and the tires and the frame of the bike, the handlebars. It all looks legit. the fact that the model again is creating all of this with code. It's just writing text down to files and then conjuring this through the computer, but it knows where to space everything out to have it all make sense and create this single uniform object. Zack Dive isn't the only one showing this off. Reuben here vibecoded this very rudimentary 3D printer. It's honestly kind of detailed. You can see the wireframe there showing off all the different portions. It's accounting for everything really. And again, it's making this 3D model has to understand the location, the distance, the sizes of all of this to make it proportionally correct and make sense as a singular 3D object that really exists in the code space. Daniel showing off a pneumatic cylinder simulator. It's labeled that has a couple different views. Volutric air flow dynamics, a live HUD that shows pressure, volume, velocity, and force. Here it all is. I just don't know what prompts I suppose they're using. Or maybe it's the interfaces, the aenic interfaces that they are testing this model out inside of that are able to command and leverage the model this way. I'll show you some of the results that I ran into testing 3D things. DD shows off a 3D ocean simulation. Pretty decent wave dynamics, reflects the sun. As DD points out, 3D graphics is extremely hard. This is still a simple demo, right? It's not something you would use maybe in a real video game, but it still has to encode 9 to 10 techniques based in physics. Being off by one number can totally screw up your results. Yet, this is now getting to the point where it's pretty trivial for an LM to produce a result like this one shot. Slap a simple prompt into the interface and works first try. Daniel showing off one that absolutely blew my mind. This is full double wishbone suspension designed simulated. I actually attempted to recreate this one myself with Gemini 3. 1 Pro, but this is just mind-numbing to me. You can see the spring and the actual suspension, the double wishbone right on top here connected to what would be the actual frame of the car. You can see it's even visualize the line that leads down to the axles and eventually steers the car. Plus, you can also see the brake line here running over to the calipers. The wire view even showing that they are three piston in design and even simulates going over bumps and the suspension moving as it drives. This is so cool to me and something I've tried to show off and do before with AI coding and I haven't been able to achieve it yet and this guy just is nailing it first try with 3. 1 Pro.

So, inspired by that, I took it a step further. I'm having the AI compare two different generations of similar vehicles talking about engine suspension and a few other things. There's graphs. It has to build a website. It also has to do 3D. So, it's no small task in just the basic, you know, Gemini web interface, the chat interface. It feels much lazier even with a good prompt. You have to demand extended thinking. And even then, sometimes the web app is just going to fail you. Regardless, it's still built something that works first try. You can see we're looking at 400 lines of code. No, it's not super long or anything like that, but it gets a lot done for how big it is. All right, so the page it built. Pretty cool title here. It's trying to show off and simulate the different types of suspensions. You can see there's a slider for suspension load, and it actually is doing something back there. You can see there is the wheel itself, and it moves if I slide this around, but it's just messed up. It's not nearly as detailed as what you just saw. And it's like in the background. I can't actually see it or interact with it. Then it compares mass and momentum. This part looks pretty great. It's very well aligned. The overall website feels clean and modern, showing off the different torque curves and comparing to the old generation. But then that's it. It kind of just stops there. It's like it didn't finish the project. It ignored a few things in the prompt and in the general scope because it spent so much time trying to code the 3D in the background. And honestly, I think the thinking budget is just super limited in the Gemini web app itself. Maybe it can dynamically be changed and if you request it in the prompt and stuff, but you really can't trust it for massive projects like this. Not unless you switch it to Deepthink. I attempted to

eliminate the Gemini web app situation by running it right through the API. Yeah, Gemini 3. 1 Pro preview is already available in the API. Catch up OpenAI. Where is GPT 5. 3? Anyways, forcing it on high thinking I think definitely made it work a little bit harder. You can see as I scroll through all of the different thought summaries that it had as it was working, goes through the whole deal and then writes the final HTML. And if I open it up, here is what we are greeted with. Again, very nice title and name. It's starting to get really great at just web design and nailing all of that first try. You can see there's a little animation scroll to initiate breaks down this mass paradox. I can't help but think maybe there's supposed to be something here on the other side, but it just kind of feels empty. Feels like again it needs another pass. It needs to sit down and spend some more time refining this. This is supposed to be a graph with a bunch of different torque curves for the various engines in the vehicles that it's supposed to be comparing. But as you can see, it's broken and not working. Here is the 3D suspension comparison. Also broken and not exactly working. But what happened? Well, we only had one pass of this model, even on high thinking, right through the API. Sometimes it's not just going to nail stuff like that perfectly.

The Google API, the AI studio, which you can access online, has a vibe coding interface that allows you to build apps. I was like, "Oh, Gemini 3. 1 is available through this. Let's send it on through. " And it is still working. As you can see, about 8,500 seconds, and you can see it's actually been building stuff just like an Aentic IDE, anti-gravity or something. It's made all the folders with the index html file and a few other things. I don't know if it errored out or something like that, but it was on its way to something. Gemini Deep Think, which just had an update. I don't know if this is technically running off of Gemini 3. 1 Pro, but I put it up to the same task. In Gemini Deep Think, it did a great job. forcing it into that heavy thinking mode to then give you one HTML file definitely gives you a better result than just 3. 1 Pro inside of the web app where the API will. It even has its own custom scroll bar. And as we scroll down, we can see the different generations of the vehicles we requested a look at. Gives you all of the different data you need comparing weights on a singular graph. And it's actually a very nice looking graph. There's little highlights around the numbers colored differently. We've also got another graph. Again, same colors, a neon vibe to it. Accurate graphs with torque curves. Very, very nice. And then you get down to where it tries to simulate the suspension. At least we can see something in the 3D space. This is the first time I got, you know, like a real interface showing off the 3D and then the manipulation on the side. Double wishbone suspension. Okay. Not really as good. but not looking as good as the other double wishbone example we saw earlier from the community. I think that's supposed to be the wheel. And you can see it's spinning around like a token. Very, very rudimentary and basic. It's like it's unfinished or just completely broken. Something's not working right or the model just wasn't given enough token generation budget or maybe it just aired out and the code exists and we just can't see it. You're supposed to be able to do, you know, airborne or compression not really working or doing any good.

So, finally, I tried this inside of Google's anti-gravity, and this one I allowed it to do multiple passes. Yeah, Gemini 3. 1 Pro is already available for free inside of Google Anti-gravity to try. I recommend you download this. It's the best for building apps right on your computer. But really, it's capable of so much more than that. Anyways, this one took more like 15 minutes. It was able to go back and forth, actually access and see the web app it was working on. And honestly, I think this unequivocally delivered the best Gemini 3. 1 result. For this one, I also requested that it go with a skuamorphic design just to change it up a little bit. And as I scroll down, you can see it definitely tried to adhere to that. I think this text would look a little bit better if it was white to match the rest of it. This is very iOS 6 feeling. Lots of shading going on. This part kind of feels indented where this sticks out a little bit more. Easy enough though to read all of the information about the cars. And what's cool is that there are some fancy animations as you scroll down. Like once we get to weight gain and the torque revolution shows a couple of different things. We got these red bars. Again, I love that little shading. It's all about those smaller fine details that you don't necessarily recognize right off the bat. And I think this is actually supposed to look like stitching. Yeah, it all looks really quite good. The skuomorphism, it's heavily documented, but definitely requires some more precision and detail to design. Makes it easy to understand the different values. Here we've got some 3D visualizations of the engines. Trying to show off, you know, a turbo engine versus the naturally aspirated ones. Extremely rudimentary, but it did its best. There's some like basic lighting going on. It's okay. Hey, we can see like the pistons at the top and the firing. I like it. At least it is interactive. As I move this dial down here, the engines start to get a little bit more crazy and more violent. They're moving faster and faster. And you can see uh the donut on the right hand side here. This is supposed to be a turbocharger for this engine. It is glowing red hot because it's getting warm and the lighting is reflected in the scene, which I quite like. It also attempted to do the suspension and it just again kind of fails. We got spinning tokens instead of wheels. At least they're shadows. You're supposed to be able to compress the suspension. I like the spring over here on the side with the McFersonson strut, but I can't really see it all the way. It tried its best. Now you are seeing how impressive those other results were and how much the prompting and the interface where your language model is actually working and how it's working matters so much and can completely change the results of the outputs. But yeah, this 3D stuff, it's new. We're starting to get into this territory with these AI models, but it's certainly not something that you can just generate with on the fly super easily. Voxels, you'll probably have a better chance. But yeah, especially interactive stuff, it's just going to be difficult. I don't plan on comparing Gemini 3. 1 to ChatGpt 5. 2. Gemini 3. 1 is honestly better. Just stick with that. ChatGpt Pro and Extended Thinking are good, but we're waiting for 5. 3. I hear Codeex is great. Regardless, I would

like to compare Gemini 3. 1 Pro to Claude Sonnet 4. 6 and Opus 4. 6. So again, this is the same exact prompt. This is sonnet 4. 5 without any extended thinking. It still took this project very seriously. And you know, that's just something about the Claude models and how they work with even the Claude web app or the co-work/code interface that you can download for your computer. It feels more agentic. It feels more personable and alive whilst also being pretty much just as capable, but likely more expensive in the API. Sonnet 4. 6 6 wrote 2400 lines and upon opening it, honestly, man, I'm already impressed. A much more animated interface and it looks like it's ready to compare. This is like a living, breathing slide almost. Scrolling down, you can see we're already getting some direct comparisons and we're even getting some visuals for the vehicles showing off how big they are compared to one another. Although I don't know if it's working out so well. It still seems to be kind of struggling a little bit. But yeah, overall comparing these pretty well. This green scan line, I got to say it's super annoying. Then showing the other vehicles. Again, very, very similar. Showing off how much larger physically the vehicle has gotten compared to the predecessor. These visual guides, it's stuff like that takes it to the next level that impresses me. Gemini 3. 1, it just doesn't seem to want to do that automatically. Got a couple of graphs showing off the bloat in terms of weight, how cars have gotten heavier over time. There's a little bit of clunkiness, I think, to the design that shows off Sonnet's size a little bit more, but man, this is still so impressive. Look at this graph curve. I think it might have generated this. These graphs look a little bit wier than they are supposed to be, but otherwise pretty solid. It's cool to see them all compared like this. It's trying to use onion skin or opacity to show everything off at once, but it's not working out amazingly. All right, now we get into the suspension DNA. It's doing 3D. Yes, sit 2. 6. This is a very rudimentary suspension layout, but it's pretty complicated. It's pretty indepth. The wheel, I mean, it's not spinning like a token, which is good. and the double wishbone. I can see both double wishbones and where, you know, the actual strut would be. And there's the dial. I can actually, you know, compress the suspension. The wheel is kind of moving a lot, but it's trying. It's trying. And then the rest of them being McFersonson, the same type of suspension are all very, very similar, but it's sort of the same deal. At least, you know, you're getting a good visual representation of the different kinds of suspensions. It's able to compare them accurately. And then you get your final analysis, which I really like. Showing off power gain, weight growth, transmissions, driver engagement, track capability. It's all very nice. Honestly, I think this is a better result maybe even than Gemini 3. 1 Pros result. And this was just through the clawed web interface. It feels like they set it up to be agentic and hardworking and take things to the next level right inside of the web app through Claude. Although their rate limits are tiny compared to Google's, this result is like pretty much as good as what we got from the anti-gravity result. Our best Gemini 3. 1. Not to mention there is an Opus 4. 6 and I ran this with extended thinking. And this website right here really incredible. Again, you'll notice it's honestly kind of similar to the one Sonnet 4. 6 six produced even down to the different colors. Regardless, we get all of the comparisons. Again, it's showing the little cars off. I honestly would like to see the size comparison more. These are just sort of visual cars. Here we get a basic suspension visualization, but this one is actually very functional. It helps you learn about the different types of suspensions and how they work. And it even covers the rear suspension. Here is the McFersonson strut. Yeah, it's not fully 3D, but I think as a learning tool, this one shows off the compression much better. You can see the wheel move. line showing how the camber is changing. And where with double wishbone, that doesn't happen nearly as much. Also demonstrating how multi-link rear works. This is the type of rear suspension all the newer cars use. But a great visualization of the compressed spring and the multiple links. Also showing off the old torsion beam design and how it kind of connects the rigidity and how if you hit a bump on one side, it affects the other side of the car. We've got little blurbs and writeups, bar graphs comparing all of the weight. Very, very nice. The torque curve graphs also looking really good. These are probably the best ones I've seen out of the bunch. Especially like the Honda Civic SI one here, showing off, you know, that quick early turbo engine making more power. You can see this one's going much farther in the rev range. Super cool stuff. And then gives you a final summary. My final test is a

synthesis that combines multiple disciplines. Xenobiology, astrophysics, atmospheric dynamics. We're talking about materials engineering. It's an attempt to synthesize what other forms of life might look like on known existing planets that inhabit the quote unquote Goldilocks zones you might see on certain star systems, Earthlike planets. you know, what would life actually look like on those planets out in the stars? So, not only does it have to do all of that accurately, but it also has to present it in an engaging, creative, substance driven way that is going to immerse us. Here is Gemini 3. 1 Pro's response. You can see it kind of gives us a little bit of a summary. Starts off with a nice big title project. I like that. Explains how it approached things. We get a spreadsheet style breakdown. So, this is the planet it chose to go with. It talks about the star that it surrounds. Illumination is deeply shifted into the infrared spectrum for this particular star. That is a biological and colonial implication. On this planet, one side has eternal day and then the other side is internal night. So like half of it is flaming hot and half of it's frozen. Life happens in the terminator zone. The ring of perpetual twilight. Talks about the atmospheric pressure gradients, relentless hurricane forces, prevailing winds. This planet's about 93% of Earth's gravity and there is ionizing radiation and UV spikes. It starts to dive into the florora. Chlorophyll as we know it on Earth is useless here. So plants can't make energy as they typically would. You would see pitch black or deep iridescent purple plants absorb maximum thermal and light energy. Fun fact, the idea of purple planets that support life is not actually new. So in a kind of strange way that tracks there would be no tall trees. Terminator winds would snap them. Apparently the flora would be crusts sprawling flat biological mats that hug the silicut bedrock of the planet. Aggressive fractal root systems that tap into deep hydrothermal reservoirs. Dualpurpose hydration and geothermal energy extraction during stellar eclipses or flare induced atmospheric darkening. Animal equivalents on this planet do not possess eyes. Vision here is entirely thermal. They navigate via highly evolved pit organs, mapping the world in gradients of heat. To survive the winds, apex grazers are armored, flattened organisms somewhere between horseshoe crab and manta ray. They don't walk. They lock themselves to the ground using powerful muscular suction, slowly grazing on the black radio mats. It then goes off to describe what it would be like if humans tried to inhabit here. If you guys want to pause and take a read, feel free. This is the final immersive experience. If you are a person visiting this planet, the son of Trappist one is permanently fixed on the western horizon. It's a dull red amber that never moves. The sky is also red. He would need a haptic suit. A faint iridescent purple sheen under the red starlight is what those plants on the ground would look like as you're walking around. A colony or a set of buildings built on this planet would have to be built like silver teardrops so the wind can simply pass over them. Pretty awesome stuff. It feels accurate. Although I am not a professional in any of these fields, so I don't know, nor can you really even prove any of this correct. But as a thought experiment, you know, I also tossed it right into Claude 4. 6. 6 sonnet. Again, this one just behaves more like an agent. I'm not saying that Gemini 3. 1 Pro didn't search the web, but if it did, it didn't bother showing or telling me any of the sources. Here we can see Claude went and searched the web for these results several different times, going through collecting the data it needs to then produce its version, which I think is a little bit more creative. It has source receipts up front as promised. This is adhering, I think, a little bit better to what my prompt asked, explaining the different planets. It actually picked the same planet and then immediately jumps into the immersive experience that I asked it for. I think the structure is a little bit better here. So, yeah, we go through this creative narrative experience. It's very detailed. It's very long, honestly. But this is the level that I asked for. First principles, what life on Trappist 1e would actually need. breaking down the basics for us before it goes into the explanation. Google 3. 1 Pro did not do that. It's actually taking us into deep biomes. Biome a burning garden. Claude 4. 6 sonnet claims life cannot rely on sunlight in any form recognizable to a botonist. Instead, the first rung of the food chain is occupied by what we might call ritoes. And just like Google, same conclusion. Matt forming chemoliththotroofes analogous to Earth's iron sulfur bacteria but far more complex after billions of years of evolution grow in sheets across mineral encrusted geothermal venting fields drawing energy yeah dense dark almost crystalline just like Google sonnet 4. 6 Six notes that the wind is a significant issue. Any creature that tries to walk upright gets sand blasted. Low and heavy works. Functionally convergent with Earth's limpids. Creatures that hug the ground with wide flat bodies. Yeah. Crustaceans anchored by powerful muscular or adhesive appendages. Dorsal surface is armored. Mineralized exoskeleton silicon and iron from the geochemical soup they inhabit. Sonnet 4. 6 also knows that vision is not going to work here. It's all about pressure and chemistry. Cheoceptive arrays sensitive to sulfur compound gradients, pet shifts, and subtle electrical fields. Living tongues tasting the landscape rather than seeing it. Navigating the hydrothermal landscape by following the gradient of warmth and chemistry like a hound follows a scent. The closest thing that they would have to intelligence, this is beyond the scope of Gemini 3. 1. Now, distributed intelligence of an octopus writ large. Decentralized nervous processing, environmental sensing across the entire body surface, problem solving oriented around chemical mapping and material extraction rather than abstract symbol manipulation like we have here on Earth. They wouldn't build cities. They would be cities in a sense. Cooperative superorganism colonies. Very interesting. And that's just one biome, guys. Next is biome B. Glacial ocean. Wow, we got biome C as well, part four, human colonization, the different challenges that faces. I mean, guys, talking about the star, and then, you know, a suggested next angle, this is a way, way deeper response than I got from Gemini 3. 1 Pro. I'm not going to lie. And I also tried this with Opus 4. 6 extended, and it wrote me a whole HTML file instead. I didn't even ask it to do that. The issue is that this used up 90% of the session limit and terrible rate limits compared to what you get with Google. This right here, this is on another level. It's even got some sort of a quote here. It starts with all of the evidence right up front, receipts backed, and then it breaks into it like it's presenting everything through a website experience. Kind of explains the premise here, what we're talking about isn't prophecy, grounded extrapolation, and it breaks it down into five different worlds. So, it's covering it on a scope 5x probably what we just saw from sonnet. Oh, I like this. I like the design here. Kind of giving us all of the basic planetary stats. Paragraph level rundown, speculative xenobiology, and then the colony scenario. Okay, so each one of these is a little bit less detailed than what we just saw from Sonnet, but it's getting pretty close to what Gemini 3. 1 Pro is giving us. This is so crazy. And yeah, like I said, I'm sure that Gemini 3. 1 Pro is capable of something like this, but you need the right prompt. You have to kind of force it and really steer it into that direction. It's certainly not going to do it from the Gemini web interface. Absolutely insane. So, after all of

that, conclusion time. The truth of the matter is that through the API, if we compare Claude Sonnet 4. 6 to Gemini 3. 1 Pro, Sonnet 4. 6 six is actually more expensive. $3 per million input tokens and $15 per million output tokens versus Gemini 3. 1 Pro, $2 per million tokens input and $12 per million tokens output. But as you just saw, it's pretty easy to get a comparison result where Sonnet 4. 6 is actually outperforming Gemini 3. 1 Pro with Clawed Opus 4. 6 being even better but being even more expensive. If you're building projects, if you're doing vibe coding, both of these models are going to be super capable. But yeah, Gemini 3. 1 Pro is super solid. But if you need the best of the best, those quad models seem to be where it's at right now. The video input recognition and overall image recognition on Gemini 3. 1 does seem to be second to none, and it's probably the best bang forbuck model out there. But I think compared to Claude, it's a little bit lazy. you need to prompt it very well and very specifically and demand extended thinking. But fun stuff, guys. Let me know what you think down in the comments below. If you've tried these models, what is your current daily driver? What is your scenario? And how do you get an AILM to attack it? Tons of stuff going on. I love to see the competition. 3. 1 Pro is a good model. I'm happy about this. Even though Sonnet is still going to be what I prefer, the rate limits are gone like that. And then yeah, I'm on to Gemini 3. 1 Pro. It's fast. It's good. Google, keep up the great work, but please unlazy your models a little bit. I don't know what it is about them, but they like to be more concise. Thanks so much for watching. I'll see you guys in the next video, and goodbye.