Google's Gemini 4 Argon AI Just Landed

Julian Goldie (0:00)

Google DeepMind have announced Gemini 4 Argon, which is their new Frontier model. Now, this is designed for complex workflows, across coding, enterprise knowledge, as you can see right here. You can see how it performs on the benchmarks. So it's pretty much crushing on the benchmarks versus GP26 Astra, Fable 5.1, Opus 5.5, winning on all the benchmarks. Obviously, you want to just test those out yourself. You don't just believe everything that you read about benchmarks, but that is a positive sign. Bear in mind, like Google, just haven't released anything for ages. But when it comes to AI, so when they do usually release a new model, it is massive. It is massive. So let's see what we get back. And also, this has a 1 million token output limit. So what that allows it to do is basically have a deeper level of reasoning for long multi-step problems in one go. You can actually find out more on the page right here. But this is the announcement of Argon. So this is their next frontier of intelligence. They seem to have totally skipped Gemini 3.5 Pro. That was supposed to be released earlier this year, but I don't think they've changed it now, so they don't want to use it basically. Now, probably the first question you're going to ask is, is this model out or how do we get access to it? So, apparently, it is rolling out right now to a set of trusted cyber defenders through the Fairwind program. So, that seems kind of similar to the Glasswing program from Claude, if you remember that, where basically they have a bunch of cybersecurity people test it first, and then they'll roll it out to everyone else eventually. So, just to be very clear here, it's not available for everyone yet. You can see in the announcement here that it says rolling out soon. They literally put this at the bottom of the article, but it's the most important thing, of course. So, here you can see they say, we built Gemini 4 Argon with frontier level capabilities, blah, blah, blah. So, don't really tell you when it's coming out. And they say, we're grateful for those people who are testing it right now.

But they don't really say when it's coming out or anything. So, it could be a week, it could be two weeks, it could be three weeks. They've made the announcement on all the benchmarks and claimed that it's better.

Honestly, by the time it's come out, it could be replaced by Sonic 5.6 or something like that, that actually outperforms it. So, that's something to be aware of as well. It's like, for the average person, we can't access this yet. It's going through that sort of testing phase. Now, let's have a look here at the launch price. So, this is an introductory price, as you can see here, of $2 per million input tokens and $10 per million output tokens. We've cashed input tokens priced at 95% off.

They've also said they're already using it internally, and some of the ideas or examples of what they use it for. So, quantum algorithmic optimization. So, basically, Argon is helping quantum computing researchers, and it beat the published baseline by 40% in a matter of minutes.

Memory efficiency, live cell code base optimizations. So, some complex stuff as well. And also, it's been optimized in some open source software as well.

One thing to note here is that we're literally talking about hypotheticals here, because none of us could test it. And so, when we're looking at these comparisons of like Gemini 4 versus Jupiter 6 Astra versus Fable 5, etc.

This will be a big model. Like, undeniably, this is going to be a really powerful model, and probably one of the best or biggest releases this year. Having said that, we can't test it right now.

We don't know how powerful it is, and any sort of guesses around this are literally just guessing. So we could have a look at the benchmarks to say, okay, is this going to be better than GPT 5.6, or better than GPT 6, or 6.1 Sol, or whatever, but all of us are just guessing. If you actually, this is pretty interesting. Because when I saw the announcement, I was like, oh, this is going to be fun. Argon's out, big release. And then I started having a look through the trending, and you see how nobody's got any outputs from it. There's not even any demos from it. It's literally just one thing over and over again, which is the benchmarks here. So it's kind of a shame, like they should demo it, or at least show some cool stuff that they've built with it. The benchmarks are exciting, but they're just numbers. We're just getting excited over numbers on a sheet. Yeah, you might also ask, like, which benchmarks actually matter for practical work? So one thing to look at is deep SWE, because a lot of us are like building or vibe coding our website or something that's working long term. And so with Gemini 4, you can see here that it is outperforming all the other models just. I mean, the percentage difference isn't massive here, but you can see between Fable 5.1, Jupiter 6 Astra, Opus, and Gemini 4 Argon. Gemini 4 Argon has the best score on long horizon SWE, so that is better. For a lot of people watching, this is not going to matter so much, the CWE.