Gemini 4 Argon

0:00 Okay, so Google has finally announced Gemini 4. So, this is actually called Gemini 4 Argon. I should preface this by saying this is not out yet. This is still in testing, but since Google's

0:12 actually announced this, I'm going to go through some of the key things that make this a really interesting model. So, first off, contrary to what a lot of people are saying, this is definitely a

0:20 frontier model in its intelligence. And I'll go through some of the benchmarks in a second and talk about that. For me, the probably most interesting thing about this is that this can actually

0:29 generate a million tokens in a single response. So that's roughly eight times what any of the other frontier models will let you output right now. But that million token ceiling there isn't the

0:41 only number I think you should be paying attention to. I think the biggest thing with this model and the number that matters most is how many tokens it actually uses to get the task done. And

0:52 really that's where this model gets interesting. We know for example a lot of the recent models including things like Gemini 3.6 flash 3.7 had issues where they're just outputting

1:05 very large amounts of tokens to actually get a task done. And that's one of the things that's actually measured by an independent party with the artificial analysis benchmarks. So I'll go into

1:15 those in a second but first let me set the context of what's actually going on here. So, Argon is the first Gemini 4 model and really Google is pitching this as a model that can do long horizon

1:27 coding, knowledge work and be a sort of key model for cyber security defense there. Currently, this is rolling out just to selected users, but you should expect broader availability in the not

1:38 tooistant future. Now, artificial analysis have already run this through their intelligence index and it scores 53 there. So that's basically the same as sort of GPT Astra at max reasoning

1:50 level and just slightly one point ahead of the GPT 6.1 Soul which was announced at dev day yesterday. But I guess in some ways for Google most importantly that this is 23 points ahead of Gemini

2:03 3.1 Pro preview which was their last Frontier Pro model. And like I mentioned in other videos, that's seven months that they didn't actually release a sort of pro or above standard model. Now,

2:15 we've seen them release multiple flash models. And to be really honest, in many ways, that's their bread and butter. Industry tends to like the much cheaper, much faster flash models that still come

2:26 with the Google and the Gemini pedigree. And I guess the other big thing with Argon is that this finally puts Google back in the top three labs on intelligence. All right, first off, let

2:38 me go back to the million token output. So, Gemini 3.8 Flash topped out at around 64K tokens out. Opus 5.5, Fable 5.1, Astra, they all cap out at 128K. So, Argon goes to a million tokens out

2:56 in one shot. Now, why does that matter? The biggest reason I think is that it just means that you stop losing things at the seam. When the model sort of hits the output cap halfway through a hard

3:06 task, your harness has to sort of stop then often summarize or compact something and then kick it off again. And every time you do anything like that, you're throwing away a lot of

3:17 detail. Now, often there's benefit for doing it in that you're saving money, etc. But as we're starting to see caching prices of models go way down, then the temptation is actually to be

3:27 able to generate more out in one shot. So the idea with a million tokens is that everything that the model has generated stays in view for a whole run. The second advantage of this is it now

3:39 means that some of the thinking tokens or the long chain of thought can actually be a lot longer in what they're generating out. And in the past, that meant that models with a lower cap were

3:50 tending to sort of rush their thinking or it was getting cut off in the generation. Now, of course, we know that that doesn't mean you're always going to get to see that thinking, but you do

3:59 hope that you'll eventually benefit from the output of that thinking. The third advantage of having the million tokens out is that simply the model can now generate much bigger artifacts in one

4:10 place. So, if you need a full module rewrite, you need some kind of migration done, you need to try and generate your Minecraft game in one pass, the model's going to be better at doing those kinds

4:21 of things because it's going to be able to retain variable names and other key bits in that pass as it goes along. Now, Google actually gave an example of this kind of work with a video decoder where

4:33 Argon agents ran many rounds of experiments to rewrite a codebase fully new in Rust to make it 2.7 times faster. Now, on the practical side, a run like this is obviously going to take a while.

4:45 Google hasn't published generation speeds for this, but if it's something like 100 tokens a second, a million tokens is going to be close on 3 hours of decoding. So perhaps that even hints

4:58 that we're going to get faster decoding from this model than we've seen from the previous pro models. We know that the 3.8 flash is doing around 300 tokens per second. And obviously with OpenAI's

5:10 announcement of ultraast, there's definitely going to be a movement for people wanting faster decoding. Another big advantage of that is that as you're getting smarter models, your harness can

5:20 actually get simpler in how it actually does things. Because if the model can do more in sort of one shot, it needs to have less loops of actually going through the model. Now, of course, there

5:31 are a few caveats here. If you're building sort of tool heavy agents, this is going to matter less than it sounds because those usually work across a lot of separate calls anyway. All right, the

5:41 next thing up comes from the artificial intelligence index. And this is not just about the raw intelligence of the model. Over the last 6 to9 months, pretty much all of the labs have been working on

5:52 shorter chains of thought that can do the same amount of intelligence or do better. Getting models to reach the same answer within a few tokens is starting to become more important than just

6:04 purely having the best model with the most amount of intelligence. And that makes sense because if you're running these models in production or for doing certain tasks, you really don't care

6:13 about how many tokens the model outputs. you care about how quickly can it actually do my task. In fact, if you think about it, you're kind of paying per task. And so the big win here is

6:24 that fewer tokens per tasks means lower cost and lower latency. And this is my one big issue with 3.6 flash was that that while the model was very smart, it just chewed up so many tokens to get to

6:38 its answers. So you can kind of think of this cost per intelligence index here is like the hidden pricing of a model. So Argon itself is $2 in and $10 out per million tokens at launch. So that's a

6:52 50% launch discount for this. And the cashed input is 95% off. And I would say that unless things dramatically change, you could think of that as probably going to be the price of this model in

7:05 that by the time that launch discount runs out, there will be newer models that are coming out to replace this. And you could think of that really as penalizing companies that stay on a

7:15 model and are not continually updating to the latest model. Anyway, at those prices, artificial analysis puts this at $1.99 per task, about 60% of Astra's $326.

7:27 And they're pretty clear that that advantage comes from the lower token price, not from using fewer tokens. If it's just purely on the number of tokens, then this model is roughly 1.2

7:38 times Astra. And I guess I should point out that while it's coming in cheaper than GPT6 Astra there, it's still about 2.7 times the cost of GPT 6.1 Soul at max. All right. Now, the other really

7:50 interesting thing the artificial analysis results show is hallucinations. So according to their eval for this, Argon has a 15% hallucination rate, Astra is at 51% and GPT 6.1 Soul is at

8:04 54%. So basically it shows that Deep Mind has done a lot of work in getting Argon to just be much more likely to tell you that it doesn't know than to make something up. Now the trade-off at

8:15 the moment in that benchmark is that actually the accuracy is lower than Astra on that benchmark. So if you look at the overall omniscience scores, you end up at roughly about the same. But I

8:26 do think for a lot of production work, it's actually much more useful to have a model that says I don't know than to be confident about a wrong answer. Another interesting thing that artificial

8:34 analysis points out here is that this model is clearly better at agentic work. It's number one on the automation bench at 77.5%. Unfortunately, on Terminal Bench 4, it's

8:44 at 507%. Which while it's still a massive jump up from 3.1 Pro, it's still a bit behind Sonnet 5.5 and Opus 5.5 and Astra there. All right, so just to keep this video

8:56 short, I'll finish off by saying Argon is a strong model and the million tokens of output opens up some genuinely new ways to think about longunning tasks. A lot of the tasks that you've seen people

9:08 doing demos with on Opus 5.5 and some of the OpenAI models, Argon is definitely at a similar level to that kind of model. And I do think a lot of people are going to enjoy using this model from

9:19 what I'm hearing is not too far off into the distant future. So when it's out and I am allowed to, I will certainly be showing some demos of this and showing you some of the testing of the really

9:28 long runs and how long they take and what cost per task, etc. that we're seeing on real world tasks. Anyway, if you want to find out more about this model, I've got links in the description

9:38 to the blog post about it. They've got Google's own benchmarks in there, which I didn't cover in this video. And let me know in the comments how you feel about this model coming out, what you think

9:46 are going to be the most interesting use cases, etc. For me personally, I think just having another company that's pushing both OpenAI and Anthropic constantly is always going to be a big

9:56 win. And even if Google's not going to be the most intelligent model out there, it's definitely good to know that they've got a model that's back in the game here. So anyway, as always, if you

10:05 found the video useful, click like and subscribe, and I will talk to you in the next video. Bye for now.