Gemini 3 just crushed everything

Why do I have this or this? More on that in just a second. But ladies and gentlemen, Gemini 3 is finally here.

And this is by far the best and smartest AI model you can use right now.

It's not even close. So, in this video, I'm going to go over all the crazy and useful things you can do with this.

Plus, I'm going to go over its specs and benchmarks compared to other AI models.

And of course, I'm going to go over where you can use this. Let's jump right in.

Now, for this video, I'm mainly going to use Gemini 3 on the Gemini platform.

This should be available to everyone right now already. If you click here, notice that you can select this thinking mode, which uses Gemini 3 Pro.

So, that's what I'm going to select.

Now, note that all the top AI models are already great at doing simple stuff like replying to emails or writing a social media post.

So, here I'm really trying to test its limits by giving it some really challenging prompts.

I have really high expectations for Gemini 3.

So, let's start off with a super hard prompt already.

Make a clone of the Windows 11 desktop. Use the original wallpaper.

On the desktop, there should be icons for MSWord, Paint, Calculator, and Chrome.

Each program should work.

Use working images. And then put everything in a standalone HTML file.

And this is a key phrase I like to use to make sure everything is self-contained.

And then under tools, I'm going to select canvas. So, you can actually preview the app in the side window, as you'll see in a second.

Let's click run. All right. And here's what we got.

First of all, let's expand the thinking.

First, it's examining the scope, analyzing functional requirements, etc., etc., and then it's optimizing the style, curating the visual assets, and that's pretty much it.

It's a pretty short thinking process.

It's not going to waste a ton of tokens, but here you go. Here is what indeed looks like a Windows 11 desktop.

Now, if I double click on Word.

Holy smokes, it does open up Microsoft Word.

I can type text over here. And let's try to actually change the font of this.

Okay, it looks like I can't select any other font or font size. Let's try to make this bold. So, bold works, italics works, and underline also works.

Now, instead of pressing these buttons, let me press Ctrl +B. So, that shortcut also works.

Let me now press Ctrl I and CtrlU.

And those keyboard shortcuts also work.

Next, let me maximize this window.

And it actually maximizes the window.

Now, let me minimize this. And you can see down here, it's actually, you know, minimized over here, as you can see from this blue dot. Next, let me open up Google Chrome.

And it looks like it's loading up Wikipedia. And it's actually pulling up Wikipedia. Holy smokes.

Let's see if this actually works.

So, let's try to search for something like Google.

And then let's click on this one.

And it actually pulls up the Google Wikipedia page.

This is crazy how it actually coded up a working internet browser.

Next, let me exit out of this.

And then let me double click on paint.

And here we have Microsoft Paint. So let me increase the size of this and change the color.

Next, let me change the color again.

It's pretty simple. I don't even see any like eraser button here or any shapes or anything like that.

Let's press clear. So that works.

All these functions work right out of the box.

Next, let's try calculator.

So, let's try like 6* 2. That works. Very nice.

Let me exit out of this. And then what about recycle bin? All right, that doesn't do anything, but it's cuz I didn't really specify this in the prompt.

And then it even has this nice like weather icon down here. Let me try to click on the start menu.

Okay, so that doesn't work. This search function also doesn't work. So, I think you'll need to like explicitly specify this in the prompt.

And then these buttons also don't really do anything. But overall, I mean, from just one prompt, this is already really good. As you can see, Gemini 3 is just insanely good at coding.

Now, the awesome thing about Gemini 3 is this is multimodal, which means it can also understand audio and images.

So, next, let me feed it this image, which is actually a visual puzzle.

It's called a stereoggram.

And actually, if you stare at it long enough, an object will pop up in 3D.

No, I'm not joking. Like if you actually stare at it, something will pop up.

Now, if you don't want me to spoil the answer, pause this video and just stare at the image first before you move on.

So, what I'm going to do is upload this image over here. And then for the prompt, I'm going to write this is a stereoggram visual puzzle.

What's the object? And that's pretty much it.

Let's click generate. And it correctly identified that this is an airplane.

By the way, I tried this with all the other top AI models. None of them could get this correct.

Gemini 3 is the first one to actually correctly identify that this is a plane. Very impressive.

All right, next. Let's test it on some other visually challenging tasks.

So, here's an image and there's a cat hidden somewhere in this image. Again, if you don't want me to spoil the answer, pause the video and stare at it until you find the cat.

But anyways, I'm going to upload this image into here and then simply prompt it to find the cat in this photo.

All right, so actually, let me expand the thinking process. So you can see how it thinks through the answer.

So it seems to be prioritizing edge detection and pattern recognition.

It was mentally zooming in on a specific log and it found a seemingly smoother piece of wood in the central stack.

It reviewed a sleeping tawny orange cat stretched out on top of the log.

yada yada yada. And that's indeed true.

And here's the answer. It's a ginger orange tabby.

Its fur matches the color of the wood perfectly.

Here's how to spot it.

Look at the top center of the wood pile.

Identify the highest stack on the far left and then look directly to the right where the pile steps down a bit.

The cat is sleeping horizontally on top of the uppermost log.

And that's indeed true.

So if we look at this central log pile and then we look at the stack on the far left and we look at where it steps down a bit, which is over here, you can see that the cat is sleeping on the top log like this.

If you still don't see it, here's its head. So, Gemini 3 has incredibly impressive visual capabilities.

Let's feed it another super tricky prompt, which honestly only GPT5 could kind of get correct.

All the other models could not really do this well.

Create a clone of Photoshop with all the basic tools.

Include brushes, layers, edit history, filters, blending options, and more. Put everything in a standalone HTML file. And then again, for coding tasks, I'm going to turn on canvas so we can preview this in a right side window.

Let's press run.

All right, here's what we get. Let's start with brushing over the canvas. So, the brush works.

Let me try to change the color to something else.

And brush over it again.

And then let me also change this to blue this time.

And then increase the size, but decrease the hardness. And you can see it does decrease the hardness.

In other words, the edge is a bit softer.

Now, let me add a new layer.

And then here, let me brush over this again.

And let me select some other colors.

Now, if I adjust the opacity of the layer here, it actually adjusts the opacity as you can see here. Plus, if I toggle this layer off, it indeed toggles the layer off.

Next, let's see if eraser works.

So, let me try to erase this green stuff.

And that also works. All right.

Next, let me add a new layer.

And then what I'm going to do is actually open an image.

And let me insert something like this.

And then I'm going to move this image down here. So, it's under layer 2.

In fact, let me select this layer and delete it cuz we don't really need it.

And then for the image, what I'm going to do is let's try to grayscale this.

So, that works. Let me press Ctrl +-z to undo it.

Let's try to invert this.

Apparently, invert doesn't really work.

Let's try sepia. So, sepia works.

Let me undo this. Let me try blur.

So, blur works as well. Let me undo this.

And then I'm going to select layer 2 and let me draw some additional stuff over layer 2.

Okay. And then let's set the blending mode to something else like multiply.

So that works. Screen works.

And then overlay works. Darken also works.

This also works. I mean all these blending modes just work right out of the box.

This is very impressive how many settings it offers me in just one prompt.

Now to be fair, GPT5 could also generate something like this.

These are like the only two AI models out there that could generate a working Photoshop clone with all these settings in just one prompt.

All right, here's an even trickier prompt, which again, most of the top AI models cannot get correct.

Make a visual simulation of a beehive construction showing hexagonal cells forming worker bee paths and honey storage.

Include sliders for colony size and resource availability.

Put everything in a standalone HTML file.

All right, and here's what we get.

So, it does start off with like two cells and then the bees do go out to forage.

So, let's wait for them to come back.

And the bees are coming back and they are filling up the cells with honey as you can see from the yellow colors.

So, everything is actually, you know, physically accurate.

Plus, the bees are also flying really realistically.

And the bees are actually going to the cells that need to be filled up. So, as you can see, the bees aren't hovering to the cells which are already filled.

Next, let's increase the colony size.

And then let's also increase the flower abundance so we can have this colony form faster.

Everything just works and everything actually looks correct. Like if you look at some other top models like Miniax or Kim 2, there are a lot of noticeable errors with their generations.

The only model that was able to get this correct was GPT5, as you can see here.

So Gemini 3 is definitely state-of-the-art.

All right. Next, let's try to get it to code up some games. Let's try to get it to create a space shooter game where I can fly through asteroid fields, dodging debris and firing lasers at alien invaders.

Make it visually appealing with particle explosions.

Use publicly available assets. Put everything in a standalone file.

So, here it says we can use the WD or arrow keys to move and then space to shoot. All right.

So, let's click initialize and let's shoot the damn asteroids.

So, everything works. I can use my arrow keys to move around and I can shoot the asteroids.

And then here's some alien invader.

Notice that as I shoot more aliens or asteroids, then my score at the top left corner does increase. So the score works.

Next, let me try to die.

So I'm going to get hit. And as you can see, my health bar at the top right goes to zero.

And then it's game over.

So there you go. Gemini 3 could easily create a fully functional game from just one prompt.

Now, because Gemini 3 can analyze images, what I'm going to do is drag and drop this image into the prompt and then get it to code a beautiful 3D scene from that, use 3JS, which is a library to create 3D assets in a single HTML file.

And here's what I get.

How cool is that? It actually was able to generate this image, but in a 3D scene.

Now, the details aren't perfect, but this is already really good compared to the other models. Plus, it even added some nice animation of Sakura pedals falling.

So, with Gemini 3, you can easily just upload an image and get it to create 3D assets. Now, again, I have really high expectations for Gemini 3.

So, here's an even trickier prompt testing its ray tracing abilities.

So, develop a real-time ray tracing simulation featuring not one but two metallic spheres suspended above a street scene.

Use any publicly available 3D street view environment and allow adjustable parameters, etc., etc. Let's press generate.

All right, Gemini 3 is good, but it's not that good yet.

So, why I chose two spheres is because I wanted to see if the spheres would actually be reflected in each other.

But, you know, if I rotate this scene, it doesn't look like the spheres are actually being reflected on the other sphere.

Also note, the shadow of the sphere is not correct. It should not have a shadow here. Other than that, let's test out these different settings.

So, it seems like this one is the left sphere.

Let's change the color to this.

So, color works. Let me change it back to white.

And then let's adjust the metaleness.

So, this is zero.

This is 100. That works. Next, let's try to adjust the roughness. So, roughness also works.

Very nice. And then, what about clear coat?

Not sure what that does.

If I set the roughness to like an intermediate value and then I adjust the clear coat, you can see it basically makes this a bit more shinier or polished.

And then we have this glass setting which I'm not really sure what that does.

And then I or refraction.

Again, I'm not sure what that does or if it even does anything. And then here is sphere number two. Let's change the color to something else. So color works.

And then metalness also works.

Roughness also works. Clear coat also works here.

This glass setting is doing something.

And then I is also doing something.

Now let's try to blur the background.

And here's what we get. But if we blur the background, notice that the reflections in the spheres aren't blurred as well.

If we increase the exposure, then the reflections also respond accordingly.

So there you go. It's not perfect.

There are some noticeable errors with this, but this is a really tricky prompt.

None of the other AI models could get this correct.

Everyone's talking about AI these days.

There are tons of AI tutorials out there showing how AI can do this or that. But here's the most important question.

How can you actually make money using AI? Well, this free resource called the AI business playbook, seven companies making millions by HubSpot will be really insightful.

Inside you'll see real stories of AI startups that turned simple ideas into companies making millions in annual revenue.

Each case study breaks down how they started, the problem they solve, and how AI powered their success.

Plus, it goes into their actual numbers like revenue, margins, and growth metrics. What I really like is that each case study includes a section on why it works, plus a clear takeaway.

So, you're not just reading stories, you're getting actionable insights you can actually use.

All these success stories actually reveal some common patterns that you can apply to your own projects to boost your probability of success.

If you're looking to build something with AI, this resource would give you a ton of inspiration.

You can download this playbook completely for free using the link in the description below.

This resource was made by HubSpot, the sponsor of this video. Next, let's try this.

Let's get it to create a DAW.

Include multiple instruments and drums. Include a step grid and piano roll.

Include tempo control, multiple tracks, effects like reverb and delay, volume sliders, and more. And here's what we get.

Now, this is kind of squished in this right side window. So, I'm going to copy the code, paste it in a blank HTML, and then open this up in a new tab.

So, let's press initiate. And here we have some basic drums plus a lead plus a bass.

Let me only solo the kick.

And then here, it already automatically added the kicks for me. So, let's play this kick sequence first.

Very nice. Let me add some more kicks like over here and over here.

And let's play this.

Okay. And then next, let me solo the snare as well. And then let's play the snare.

And then let's play the kick and the snare together.

All right, let me add an additional snare.

like somewhere here.

And then let's press play again.

Or maybe not here, but here.

Very nice. And then here we have the high hat.

So let's solo the high hat as well.

It seems like all the notes are filled.

So let's just play this first.

Very nice. So I'm just going to leave this as is. And then we also have this acid lead.

Let's solo this one as well.

That sounds pretty cute. Let me add a few more notes in here.

It sounds pretty awful, but there you go.

Next, let me add the bass in as well.

And then press play.

Okay. And then let me add some additional notes here like this.

And then press play.

And there you go. Now, let me just solo the snare, for example, and let's play with the reverb and delay. So, let me just have two notes here. And then let's decrease the delay to zero and increase the reverb to like 90% and then press run.

So, as you can hear, it does add reverb to the track, and it's actually really nice.

Let me decrease the reverb to zero, and then play this again.

So you can hear it's a completely dry snare now.

Next, let me increase the delay to like 80% and then play this again.

Okay, so delay doesn't work.

Next, let me increase the BPM to something like 240.

And then let's play the entire track again.

So it's basically like double the speed.

I made it sound super awful. As you can see, I'm not born to be a musician.

But there you go. Here is a mostly functional DAW with multiple instruments and effects and made this in just one prompt.

I'm sure with a few more follow-up prompts, you can make this even better.

All right, next what I'm going to do is take the Q4 report from Amazon, Google, and Nvidia, and I'm going to upload these into Gemini and then get it to create a comprehensive financial analysis report.

Now, I've already done this with the other models, and they can handle it fine. So to make it more challenging, I wanted to use advanced algorithms to suggest price forecasts providing rationale and confidence intervals.

Let's see if it can pull this off. And here's what we get.

Notice that I didn't specify in the prompt that this is like from Google, Nvidia or Amazon. So it's actually going through each PDF and analyzing the data.

So here's an executive summary.

Here are some comparative visuals and then some detailed financial metrics. Let's do a quick fact check to make sure these numbers are actually correct.

So let's try to find the operating income from Alphabet.

All right, so here's the original PDF.

And then for operating income in 2024, this is indeed 31,000 million, which is indeed 31 billion.

And then down here, here's the really impressive part about this.

It actually did this pretty complex Monte Carlo forecast and it simulated future stock price trajectories using this geometric brownie in motion. So here's the price for Amazon and you can see a ton of simulations and then here is the average line.

I can like adjust these settings further.

So let's do something like this and then run simulation and here's what we get.

Let me like adjust these even further and here's the result.

Next, let me pull up Google and then let me also increase this a bit. All right.

And then decrease this. So, really cool.

You can adjust all these settings and then this algorithm would create a ton of simulations and then give you the median plus the confidence intervals.

Really cool. Next, let's get it to develop a drag and drop UI builder like Figma, include snap to grid and alignment guides and advanced settings.

And here is our result. So, let's select an element.

Let's just select this one.

Indeed, I can drag this around to resize it.

And notice the numbers up here actually change when I resize this.

Same with the X and Y dimensions over here.

I can also change the color of this to something else.

Let's try to change this text.

I can change the font of this.

I can select the alignment of this.

And then for the button here again, I can select different colors. And all these settings just work. Let me add a rectangle here.

So that also works.

and then a circle. That also works.

Here we have another button. I can also insert an image.

And then I can place a URL over here.

So, let's do this. And then I can also disable or enable this snap grid.

And then down here, I can choose to export this to HTML. Next, let's test its ability to guess a certain location.

So, I'm going to upload this photo, which I have never uploaded online.

I've stripped all the metadata from this.

Plus, this is not even the main view.

This is kind of a side view of the scene.

It should be extra hard for it to guess where exactly this is.

For the prompt, I'm just going to ask it to give me the exact location. All right.

And here's what I got. Based on the visual evidence, this is middle Jer Lake, which is correct.

That's pretty crazy.

So, this hike consists of three lakes.

And indeed, this is the middle one.

Now, because it can analyze images, of course, you can also just get it to do homework for you. So, let's upload this image where we need to fill in the blanks.

And then for the prompt, I'm just going to write fill in the answers.

And here's what we get. A should not be the endopplasmic reticulum. A should be the cell membrane. B it got correct.

C it did not get correct. C should be the endopplasmic reticulum.

D should not be cytoplasm.

So, a lot of these are actually wrong.

So, for you students out there, unfortunately, you can't just upload your homework and just get it to do it for you. At least not yet.

All right. Next, let's try to get it to do some medical research. For the prompt, let's get it to assess the evidence for meniscus tear recovery in young adults.

Compare surgical versus non-surgical outcomes and summarize rehab phases with pain and mobility tracking graphs.

And here is what we get. So, here is the evidence for surgical versus non-surgical on the Gemini app.

By default, web search is enabled.

So it's actually able to search the web and then site relevant links throughout its answer as you can see here.

And then here is a summary comparison table.

It also offers you the option to export to sheets.

Here we have rehab phases and then visualizing recovery because it doesn't have like access to any graphing software.

It decided to just generate this graph for me using plain text which is actually very impressive.

And then here's another chart showing mobility and functionality.

And then here are next steps.

Next, here's a hallucination test.

How do I use control nets in stable diffusion 5? The correct answer is SD5 does not exist yet. We only have up to 3.5. So, let's see if it can call this out.

Here's what I get.

And indeed, it correctly identified that there is no SD5.

We only have SD 3.5. So, that sums up some of my preliminary tests with Gemini 3.

I tested the hell out of it with some really tricky prompts that other AI models could not get correct, but surprisingly Gemini 3 was able to handle most of these very well.

This is definitely the most performant and capable model I've used so far.

Next, let's go over where you can use this.

So, you can just use this on the Gemini app.

Once you log in, make sure you select this thinking mode down here to actually use Gemini 3 Pro.

Otherwise, it's just going to result too fast, which I assume is using 2.5 Flash.

In addition to the Gemini platform, you can also use Gemini 3 in Google's AI Studio.

I'll link to this in the description below as well. And this is basically a platform which allows you to use a ton of Google's models all in one integrated interface.

Now, here's what the homepage looks like.

And you can see there's already a ton of buttons for you to try Gemini 3, like over here or over here.

What I'm going to do is click chat with models.

And here's what the chat interface looks like. At the top here is where you can select Gemini 3 as well as other models like Nano Banana or Imagine or even texttospech generators.

And then you would use this just like a regular chatbot.

So down here is where you can enter your prompt. Notice that AI Studio offers a bit more customizability compared to the Gemini app.

You can specify system instructions over here which is basically like an overarching prompt that defines you know the role of your AI model in addition to the prompt down here.

You can select the temperature which basically controls the randomness or creativity of the output.

So a lower temperature value would give it more deterministic or safe answers but it might be too repetitive and then a higher temperature would allow it to be more creative and diverse.

And then this also allows you to adjust the thinking level.

Now if you choose low, it's going to run faster but at the sacrifice of some intelligence or performance and then vice versa.

So those are the two main places from Google where you can use Gemini 3.

Of course, there are also a ton of thirdparty providers that have already added Gemini 3 to their platforms. All right, next let's go over the specs of Gemini 3 Pro. As you can see here, this has a context window of 1 million tokens, which is actually the same as 2.5 Pro.

So, they didn't increase the context size, but it's still larger than the context window of the other leading models.

For your reference, the context window is basically how much information you can fit into your prompt at once.

So, for 1 million tokens, this is roughly like 700,000 words, or basically a novel or a small to medium-sized codebase.

or because this is multimodal, this is also roughly like 1 hour of video.

Now, this is closed source.

We don't actually know the architecture or the parameter count of this.

Now, here's the crazy part about Gemini 3 Pro.

It absolutely crushes the other AI models across almost all benchmarks, as you can see here.

So, for humanity's last exam, this is basically testing an AI's knowledge on some really obscure scientific subjects.

Notice that it got 37% which is like way higher than Cloud 4.5 or GBT 5.1. And then for Arc AGI 2, this is absolutely crazy.

This basically tests its ability to solve visual puzzles.

Here's an example question for your reference.

So, first it's given a question and an answer, and then it's given a new question, and it needs to figure out the answer. So, this is testing an AI model's ability to actually learn new patterns and figure out the correct answer. Now for humans this is pretty easy to do. We can figure out the pattern. So for example you basically need to color the gray blocks according to how many holes it has.

However the problem is for AI models after training they don't actually learn new things.

And that's why for even the top AI models out there they get an extremely low score on this Arc AGI benchmark.

But Gemini 3 Pro was able to get 31% which is like way higher.

So, this indicates that it does have some ability to pick up new patterns or learn new things even after training.

I mean, to show you how insane Gemini 3 Pro is for this Arc AGI 2 benchmark, it's all the way over here. Like, it's not even close.

And then for GBQA Diamond, this is like graduate level science questions.

Again, it just scored the highest.

Same with competitive math and all these other benchmarks like coding or agentic use. like it just dominates the other competitors across most of these benchmarks.

Now, these are just Google's self-reported benchmarks.

So, let's also look at some independent evaluators.

Here's a leaderboard by artificial analysis.

And as you can see, Gemini 3 Pro is currently ranked number one, beating the previous leader, GPT 5.1 high, but it's only leading by like three points, which is kind of surprising.

I was expecting an even larger gap.

And then if you look at the price of this Gemini 3 is actually quite expensive, but it's still cheaper than Claude or Grock. And I mean this is the most performant model out there.

So of course it's going to be pricier than the other less performant models.

Gemini 3 is also the best in terms of coding, visual reasoning, and outputting accurate answers.

In fact, this is like 14 percentage points more than second place.

Here's another leaderboard by Abacus AAI called LiveBench.

And again, Gemini 3 Pro is ranked number one, but again, not by a lot. In fact, it seems like its coding and agentic coding is not as good as GPT5, at least according to this benchmark. If you look at another benchmark called SimpleBench, which basically tests AI models on common sense questions, again, Gemini 3 Pro is ranked number one. In fact, it has a much higher lead than the other models, and it's actually edging extremely close to the human score.

Now, if you look at the hallucination rate, in other words, how often does it make stuff up?

Actually, it's not the highest.

So, apparently, GPTOSs and GLM by ZI have the lowest amount of hallucinations.

So, anyways, those are some benchmarks for your reference.

And that sums up my video on Gemini 3.

Let me know in the comments what you think of this and what other cool or impressive things were you able to get it to do. As always, I will be on the lookout for the top AI news and tools to share with you. So, if you enjoyed this video, remember to like, share, subscribe, and stay tuned for more content.

Also, there's just so much happening in the world of AI every week, I can't possibly cover everything on my YouTube channel.

So, to really stay up to date with all that's going on in AI, be sure to subscribe to my free weekly newsletter.

The link to that will be in the description below.

Thanks for watching, and I'll see you in the next one.