Google Just Caught Up (Gemini-4 Argon)

So Google's newest model, Gemini 4 Argon, is here and it's pretty big. Like we're talking GPT-6.1 SOL big. We're talking Opus 5.5 big. We're even talking big. And in true Google fashion, not only were they able to withstand the FOMO and not rush to release their new model. Come

on, even faster.

Patience is a virtue.

But they also continued their longstanding tradition of over-the-top naming conventions. So Argon is part of Google Search. It's part of their coding platform, Antigravity, and it even has its own new competitor to like Muse and Dots. So in this video, I'm going to break down the biggest changes, explain what they can actually be useful for, and then as always, get down and dirty with some hands-on demos and tests. So I know what you're probably already thinking. Didn't Google already release Gemini for Argon? Well, no, not exactly. About 2 weeks ago, they announced Gemini for Argon. They put out a blog post breaking down our next era of frontier intelligence. However, when they put out this blog post, they didn't give anybody access yet. It was an announcement of a future release. I mean, a few trusted cyber defenders got it, but you know, none of us normal folks. But what they did announce looked pretty impressive. And well, we finally have access as of today. Well, assuming you're on one of their higher-end paid plans, you have access as of today. Now, I don't want to go too deep into all of the details because I did break it down in my news video from 2 weeks ago where we talked about this announcement. But I do want to quickly recap what makes this one kind of a big deal. The main thing to understand about this Argon model is that it's really designed for the most complex problems. Like, it'll do the sort of like replacement for Google thing where you can ask it a question and it'll give you a pretty good response. But I mean, most models are pretty good at that now. This is for the sort of longer tasks where you expect it to go and spend a little more time on it. And the initial benchmarks do look really, really good. It's now the state-of-the-art model according to the VALS Index. which evaluates models on like professional tasks across areas like finance, coding, legal work, and taxes. It's state of the art on Automation Bench, state of the art on legal work. When it comes to agentic coding, for the longest time, most of us were looking at DeepSWE as the benchmark for how well it's gonna do at coding for us. And well, this one's now the new state of the art on DeepSWE. Another thing that's really interesting about this model is that the model has an output token limit of 1 million tokens. Now, you often see models that have like a 1 million token full context window, meaning that between input and output you get a million tokens, but most of that context is typically used on the input side, not the output side. This new model can output insane numbers like 1 million tokens. To put that into context is roughly 750,000 No other model is giving you a million tokens of output. Another benchmark I like to look at is Artificial Analysis, which is sort of a combination of a whole bunch of benchmarks to find the sort of objectively smartest model. And we can see that on this one, Gemini 4 Argon actually falls at 53. So it puts it on par with Fable, but not quite as good as Opus 5.5, Sonic 5.5, Or Muse's new Wonder model that also just came out this week. I also like to look at cost per task, and Gemini 4 Argon is way down here, so it is quite a bit less expensive than a lot of these models. It's quite a bit cheaper to use per task. Now, the group of people that are probably gonna most likely notice the big changes and the big improvements with this model are gonna be coders, and that's kind of how it's been going lately with most new model releases, is that The big leaps all feel like they're in coding. And I guess that really makes sense because when you think about it, if these models get really good at coding, well then they can write code to solve other problems. So it really seems like all these model makers are putting a ton of emphasis on making their models better and better and better at writing code. And this new Argon model is no exception. Like we said, it is now the top model on DeepSuite. This model being so good at coding, well, of course it's coming to their AntiGravity platform. This is basically Google's version of like OpenAI's Codex, of Claude Code, SpaceX AI, or SpaceX SI, whatever they're calling themselves now. They own Cursor now. So it's kind of like Google's version of all of those things. And as we can see, Argon is available in a low, medium, and a high mode inside of Anti-Gravity. Now, full disclosure, I did get early access to Anti-Gravity. so that I could test some things and make some demos for this video today. Now, I haven't tested this yet, but apparently Anti-Gravity's coding interface is also coming directly into the Gemini app on macOS, similar to the way you can open the ChatGPT app and you can switch between Codex and chat modes and work modes. And in Claude Code, you have both the normal Claude Chat and Claude Code. It sounds like Gemini is following suit and there's going to be anti-gravity coding, and, you know, chat interface all within a single macOS app. A few other quick notes before we get into testing: Google does specifically say that Argon will give more direct answers, it's going to use fewer clichés, and follow style instructions more reliably than previous Gemini models. Another thing they're rolling out with this release is not only the ability to chat, but if you're inside the Gemini app, you can actually switch over to Tasks, So to me, this seems very similar to if you're using the ChatGPT app and you're switching between chat and work. Chat is where you're gonna ask it questions and have more discussions. Tasks is where you're going to specifically ask it to go and do things on your behalf. Now, there was what they were calling Gemini Spark, which was their like agentic thing that would go off and do a bunch of tasks for you. And from my understanding, that's being rebranded to this task mode. So it's no longer Gemini Spark. It's now chat and tasks. Similar to the other companies, it also connects to a ton of apps, things like Instacart, OpenTable, Adobe, Shopify, you know, all of those types of tools that most of those other apps also connect to. It sounds like we're getting that inside of Gemini as well. Okay, last thing before we jump into actually showing what it's capable of, who actually gets access to this? Well, right now, if you're a Google AI Pro or Ultra subscriber, You'll get access in the Gemini app and in AI mode in search. If you want to be able to use Argon inside of Anti-Gravity, that actually requires a Google AI Ultra subscription. So the Pro subscriptions, I believe, start about $20 a month, and that'll get it in Gemini and AI mode in search. And then Ultra starts at, I believe, $100 a month is their lowest tier, and that's going to get you Argon inside of Anti-Gravity as well. It is also available in the API, and they did already announce costs earlier for API. It is $2 per million input tokens and $10 per million output tokens. So if we compare that to Claude's API, that's about half the price of Opus and about equal to the price of Sonnet 5.5. If you're looking at OpenAI's models, it's about the same price as GPT-6.1 SOL to use. Now, it is important to note though that these are introductory prices. And then at the end of November, starting in December, the prices go to $4 per million input tokens and $20 per million output tokens. So we're on par with Sonnet and GPT-6.1 for now during this introductory period, but then it actually will eventually get slightly more expensive in the API specifically. All right, enough talk. Let's actually see what this thing is capable of. So my first test, as you can probably guess, was to see how well it could generate Gary Busey as an SVG. And honestly, it is probably the best. Now, it's not showing up on my leaderboard yet because I had to run this test before it was publicly available and before I had permission from Google to actually put this on the leaderboard. So I don't know how my LLM as a judge is going to rank it, but I bet it's gonna be near the top somewhere. I also don't know the cost or tokens used 'cause I was using my demo account to generate this. I will have to go and regenerate it once it's live to get an accurate cost and amount of tokens used. But we can see it is really, really good at generating SVGs, something that weirdly all of the Google models have really been good at over the years. And this one's no different. It's, it's up there with probably the best SVG so far. And if you've been watching my channel for a while, you know, another thing I like to do is ask it to create a clone of the popular 3D game Mega Bonk. I give it the same prompt every time. I've given almost every model I've tested. I want to see how good But it is at sort of looking at another game and recreating it. And I like to test the graphics and the playability and everything like that. And this is the one-shot output that it gave me. So let's try it out here. So here's what our game looks like. There is apparently no audio on this home screen, but it gives me 6 characters to choose from, a couple stages to choose from. Okay, so we do have some audio when I click around. So let's go ahead and use the fox character and we'll use Whispering Forest and let's start bonking. All right. Okay, so I can move my camera around with my mouse. It's automatically shooting. We got automatic level-ups. I can jump. I could slide. I mean, it's decent. I wouldn't say it's the best looking game we've generated before. I still think Opus 5.5 is probably the clear winner in like design aesthetic of making a game. But I mean, it's, it, it's good. It's, it's decent. I also think GPT- Astra. I think GPT-6 Astra also was a little bit better looking than what this did, but the game mechanics came out pretty perfect the first time. I wanted to test building a little like physics engine, so I gave it this prompt here. I'm not going to read the whole prompt. Feel free to pause. But basically I asked it to build a little physics playground where I can make adjustments on the playground and see how the physics are affected. And here's what it created. We've got this little cannon here. We could change the gravity presets to Earth. Mars and Moon. We could change our launch angle here, change the speed, the gravity, wind acceleration, target distance. I mean, it all looks pretty good, and I love how I can change these sliders and everything updates in real time. We do fire, we can see it lands right where we'd expect it to. We could randomize a new target and then adjust things to get it to hit our target. And then down here we have sort of an explanation of everything and I mean, it looks pretty good. Then let's do a Mars simulator. So if I fire this on Mars, we can see it just floats for a lot longer and eventually misses our target by a lot. It also built in a light mode or dark mode. Cool. All right, so let's jump into Gemini on the web real quick and do some tests over here. So inside of Gemini, if I click this dropdown, you can see I have access to Argon. I am on an Ultra plan. If you're on a Pro or Ultra plan, you should see access to Argon here. And let's test a couple things inside of chat. So for my first prompt, I wanted to give it some scheduling details here, like 2 people, Alex and Blair, start at 9:00 AM. Both are unavailable from noon to 12:30. Each person can do only 1 task at a time. Tasks cannot be interrupted or split. Here's different tasks. And then I ask it to give me a schedule with start and finish times, earliest completion time, a concise proof, constraint by constraint check of your schedule. And we can see it broke down when each of these tasks would get done, who they'd be assigned to, the duration. Earliest possible completion time is 2:00 PM. So basically I wanted it to have these various constraints and then do the figuring out for me of when the different things could get done and who would be available to do them. And it looks like it did a great job at that kind of thing here. I wanted it to do some product research for me. So if I pay for ChatGPT Plus, Claude Pro, or Google AI Pro, does that subscription also cover using their models through an API in an app I build? Compare the 3 services. And you know, you could read the rest of the prompt here if you want to pause and take a look, but you can see it built a very nice looking side-by-side comparison. ChatGPT Plus, you do not get access to API calls. Claude Pro, no access to API calls. Google AI Pro, you get $10 a month in GenAI and Cloud Developer credits. And we can see further side-by-side comparisons on things like consumer chat surface, whether coding agent access is included and things like that. And then it goes into a service-by-service breakdown for each of the 3 tools so that you can compare which one is best for you. All right, there's another thing I want to test. So I was on a pre-briefing call with some of the Google team earlier, then they were walking through some of the stuff this model is really good at. And one of the things they said it was really good at was making 3D printable files. And I was like, all right, I got to try this. But knowing me, I want to do something weird and off the wall, probably involving Gary Busey. And yeah, I gave it the prompt, create a 3D bust of this person's head. And well, I gave it an image of Gary Busey. It should be a 3D printable STL file that I can send to my Bambu Lab printer and easily print out. All right. So your 3D printable portrait bust STL has been generated and verified. Interestingly, I didn't say it was Gary Busey because I was like, Is it gonna refuse to do it if I tell it it's Gary Busey? But it knew. It named it Gary Busey Bust. So, so let's go ahead and open this with Bambu Studio, my 3D printer slicer. Jeez. There we go. Look at that. Oh man, that is, that is actually pretty amazing. Uh, I don't know about that profile, but straight on, it looks pretty good straight on. Amazing. Oh, and in case you're wondering, Yeah, I actually did print my Busey bust. It looks good. I mean, I'm, I'm happy with that. I think it deserves a permanent place in my background somewhere. So I agree, it's pretty good at making 3D models that you can actually print. So I also wanted to test the task feature inside of the new Gemini, but I don't know if my demo version that I'm testing, it's not like fully working yet because I have not gotten this to work. Like this to me seems like a fairly easy prompt to be able to test. Research 5 most significant AI product announcements over the last 7 days. And it's been working on this for like an hour. When I click on plan steps up here at the top, it says no plan available yet. So it just seems to have stalled here. Now, I also gave it this task, create a 6-slide Google Slide presentation explaining prompt injection to a non-technical small business owner. And well, it's doing the same thing. You can see it's starting to think through, but it's been stuck here with the last thing it completed was incorporating real-world speaker notes. And this has also been stuck for like an hour. It just sort of got hung up at this and it's not getting anywhere. Hopefully by the time you're watching the video, this task feature has been figured out and is working. But I think my version that they gave me early access to may still have some, some bugs or something. However, When I click on Plan Steps up here in the top right, you can see that it actually did create a slide deck. I've got a 6-slide slide deck here that explains prompt injection. So it's done the thing, but it still, for whatever reason, got stalled out on the prompt. But because I'm having issues getting this task feature to work, I'm just going to leave it there. Now, there is one last cool thing that I do want to show. They've actually been using this model for robotics. And it's been able to do some pretty cool stuff. They actually used Gemini Argon to teach a robot how to play cornhole. You can see the prompt. We want to show you how we play cornhole. Can you take a video from your left eye cam? And you can see the robot here is watching this person throw the cornhole bags from their own eye's perspective. And then the robot went and practiced itself, sort of learned what it was doing. And eventually, after enough demonstrations and enough practice, was able to actually figure out the exact way to throw cornhole bags. And then it eventually managed to throw it in the hole perfectly 7 times in a row. So after watching some videos and learning and practicing itself, it learned to pretty much nail it every time. I mean, albeit from like 2 feet away, but it did figure out how to play cornhole by watching videos and taking prompts. All using this new Argon model, which I also thought was pretty cool. But that's the new Gemini 4 Argon, basically a new state-of-the-art model out of Google. So Google has finally caught up. There's been a lot of discussion around like, how did Google drop the ball so hard while Anthropic and while OpenAI just kept on pushing out bigger and bigger and better models? And well, it turns out Google was working on these models behind the scenes. They come out and they drop a model pretty much as good as those other models that are out there. For roughly the same price. Now, if we're being totally honest, do we feel like this is something groundbreaking that we haven't been able to do already? Well, outside of that million output tokens that I mentioned earlier, I don't really think this is going to like change anybody's life. I think we're getting something out of Google that's on par with what we've been getting out of Anthropic and what we've been getting out of OpenAI. But it does show that Google is behind the scenes building this stuff. And actually still on par with those other models. Now, is it likely that Anthropic and OpenAI are going to drop new models within the next coming weeks or months that just blow this one out of the water? Also quite possible. But since this was a model that I did get slight early access to and had had a chance to play with it a little bit, I figured I'd shoot a video, show you a little bit about what I've learned, and give you a quick breakdown of what it is and, you know, why you might care. I think that million token output window is a pretty significant leap. I think that's one of the better things it has going for it. But as far as like coding models go, it doesn't totally feel like a huge leap to me over what we have from Anthropic and OpenAI. But I mean, Gemini is still in the game. And one reason I never ever count Google out is because Google owns such a huge ecosystem. I mean, everybody's got Gmail, everybody's got Google Calendar. Everybody's using Chrome. Like, Google has the ecosystem. So when they come out with a really, really good model and they can just plug it into that entire ecosystem, well, overnight Google could now claim the most amount of people are using their models again, which is one of the reasons you can never completely count Google out. I think Google has a lot going for them. And, uh, now they've got a model that's just about as good as anybody else's models, which is Pretty cool. Anyway, that's what I got for you today. Every week I keep up with all of the latest AI news. I keep track of all the latest models, all the launches, all the releases, all the drama, pretty much everything going on in the AI world. I drink from that fire hose. I keep myself overwhelmed and in the loop all week. And then every single Friday I make a video that breaks down all of the news. So if you just want to keep up with one video a week to learn all of the AI news, that's my Friday videos. I highly recommend liking this video just because, you know, that'd be really a cool thing to do, but also subscribing to this channel to make sure you see those Friday breakdown videos. And I will keep you fully looped in. I gotcha. So again, consider liking this video, subscribing to this channel, and I will hook you up. Thank you so much for tuning in. I really, really appreciate you hanging out and nerding out with me today. And, uh, yeah, hopefully I'll see you in the next one. Bye-bye. Bye-bye.