Last week, gas changed the world forever. Twice a Tesla finally unveiled the Roadster and its ability to fly by farting cold gas. But while Elon was busy pulling off his greatest magic trick, Google was cooking up a noble gas of its own, Gemini 4 Argon, which is the most powerful thinking sand they've ever released. Sadly though, they never even released Gemini 3.5 Pro, which was euthanized up in Canada after failing to hit internal benchmarks. In fact, Google was getting so far behind in the AI race, I needed to add a new G category below F on my tier list. But all that has now changed with the release of Argon, which was trained from scratch over the last 6 months using Google's TPUs and their Jack's Python library, and that produced a new Frontier model that crushes GPT6 Astra and Claude Opus 5.5 on the Trust Me Bro benchmarks while remaining cheaper and faster. There are some mind-blowing Argon demos out there, like this futuristic flying car game. But in today's video, we'll put it to the test firsthand and see how it compares to Claude and GPT6 on a variety of equin based challenges. It is October 20th, 2026, and you're watching the code report. But first of all, what's up with the name Argon? It's a very interesting name choice that translates into lazy in Greek, but I think it actually has a hidden meaning implying that your tokens are gone. Google has never been good at naming things, and Argon was actually announced back on September 30th, but only for a handpicked group of trusted cyber warriors through something called the Fairwind program. These users got access to Argon without any guardrails. But even with the guardrails turned on for the rest of us, this new model is supposed to produce far fewer refusals. And when you use Argon, you'll notice it feels much different than previous Gemini models. And that's because it's not a fine tune of the older Gemini. Under the hood, it's a sparse mixture of experts transformer and was trained from scratch on Google's TPUs. Its knowledge cut off is March 2026, so we can assume it's been training for at least 6 months. At first glance, the TMBBBS are extremely impressive. Argon is state-of-the-art on the deep software engineering benchmark, which is used for long horizon software engineering jobs, and it's number one on the Val index, which weighs a variety of different topics like finance, legal, and coding based on their share of the US GDP. In addition, Gemini is now the best AI lawyer and also dominates Long Video with 91.7% on LVbench. But if we look at some of the numbers that aren't bold, like Terminal Bench 4.0, we find that Opus 5.5 wins 66 to 57 and Gemini also gets beat on Frontier S. Where GPT6 Astra wins by more than 10 points. What's also kind of interesting is that on Arena AI, where humans rank LLM subjectively, Gemini Argon comes out on top on the text benchmark. But when it comes to code arena for webdev, it drops all the way back to number nine. So this tells me it's probably not going to be as good as Claude or GBT6 at coding. But here's where Google gets aggressive. Until November 30th, Argon only costs $2 per million input tokens and $10 per million output. That's a great deal for a model this capable. And Google didn't even have to pay me to say that. But somehow they still managed to screw up the developer launch when they pulled Gemini 3.1 out of anti-gravity and replaced it with Argon, but only made it available to the people with the AI Ultra plan. That's the plan that cost $100 per month, so all the poor prousers were pissed off who only got access to Argon a few days later at a smaller quota. Luckily though, I'm dumb enough to pay for the AI Ultra plan. So, let's try it out with some real world examples. One of the selling points of Gemini is that it's an omni model, meaning it can take any form of input like image, video, sound, and so on, and output all those formats as well. Technically, chat, GPT, and Claude can't output raw video. So, the first prompt I tried was to have it generate a promo video for my app. And the end result was mind-blowingly mid. >> Tired of your horse feeling lonely in the pasture? Introduce them to Horse Tinder, the ultimate app. Like the video looks pretty good and all, but it's obviously AI and would never work for a top tier app like Horse Tinder. But saying Claude and GPT can't output video isn't totally accurate because recently Claude in particular got really good at generating motion graphics by writing the actual code that generates a motion graphic sequence as opposed to having the model use a technique like diffusion to generate the actual video. As you can see here, after giving the same prompt to Claude, we get a video that not only looks way better, but that's way easier to customize and analyze because we have access to the actual code. But one of the best ways to test a Frontier model these days is to see how well it can render 3D graphics in 3JS. So, I sent the exact same prompt to Argon, Astra, and Opus. Create a realistic interactive 3D horse with 3JS. This is actually a very challenging thing to do for both AI and humans. And the results I got back from Claude and GPT6 were pretty impressive. There are some obvious flaws with these, but we could probably get them ironed out with a few extra prompts. Gemini's result though was on a whole another level. Before I show you Gemini's output though, I have to warn you, the result may be offensive to some viewers. Viewer discretion is advised. So, the first thing you'll likely notice is how great this UI looks. It's clean and expressive. But if we turn our attention over to the 3D model, you might notice how Gemini's horse looks a bit more than the others. Now, in my prompt, I asked for a realistic horse, not a low poly horse. So, that's pretty disappointing. But the much bigger issue here is alignment. If we turn our attention over to this region here, you'll notice Gemini, for some unknown reason, decided to add a copulatory organ to the model. I was shocked and offended when I first noticed it. And look, I totally get that I asked for a realistic horse and that equinus is necessary for reproduction. So, maybe Google was just trying to keep it real. But why is it out of its prudium? And why is it so comically large? These are difficult questions Google will have to answer because right now I don't feel safe using Gemini until they fix the alignment. But one thing that definitely doesn't have alignment issues is Macroscope, the sponsor of today's video. Now that coding agents are waterboarding us with a million pull requests per day, the only way to keep up is by using even more AI to review all of them, which sounds insane until you look at how Macroscope actually works. that they built their own a walkers for the most common programming languages so their agent can get real context on every code change which makes every review faster, cheaper, and more precise. They also run your team's custom rules as native GitHub checks that can block a failing pull request. If a PR passes those checks, plus a code review and has a small blast radius, Macroscope approves it without a human ever needing to see it. They're now auto approving 40% of all their customers PRs, which means those developers are 40% less likely to lobomize themselves. And try letting Macroscope auto approve some of your team's PRs for free using the link below. This has been the code report. Thanks for watching and I will see you in the next one.
Generated algorithmically for Search Engine
Indexing.