Gemini 4 Argon vs Claude Opus 5.5: Benchmarks, Pricing, Real Tests

0:00Finally, Google's Gemini Argon 4 just dropped the last few days, and Google has been way behind when it comes to releasing new models. Everyone was waiting with anticipation while Opus and Sonnet and Astra all released in these last few months, and we kind of forgot that Google was even in the race. So, you can get it through Open Router or you can pay with a subscription with Google AI Studio, but I will have to say the results and the feedback are not that good. And in this video, I'm going to be breaking it down all on how we're going to use it, what we're going to build, how we're going to test it, and if it's even worth using or wasting your time with. You can see here Theo just released this post today on X and he

0:45says, "Gemini 4 Argon has proven to be outright unusable for real code work in my code bases. It aggressively hallucinates about how systems work." So you can see here which model understands T3 codes architecture best. You can see Opus 5.5, Sonnet 5.5, Soul Astra and so forth. Even Muse Wonder scoring at a seven and you can see Gemini 4 Argon is at a 3.9. So this is a problem with benchmarks. It's just benchmarks can be benchmaxed and you don't know if they're legit or not and if it's even worth wasting your time. That's why I make these videos to see if it's even worth testing on your codebase or in production. I think that these model drops have been so hyped up that we

1:31really need to see some realworld use cases. So you can see here I have just a basic outline on why you should possibly use it now before the price goes up and why I would kind of hold off especially if you have a claude code subscription. So what is Gemini for Argon? It's their latest Google's frontier model just announced on September 30th um for cyber security and then out to the public on like I said yesterday October 13th. So, it's number one on the Vows indexed, number one in the text arena, and 1 million tokens in and out, which is adding up to be amazing so far, but it's not number one at everything. What it is, it was announced September 30th, like I said, for Cyber Defenders. You can see October 13th, which was just

2:17yesterday, out for everyone in the Gemini app, Pro and Ultra subscribers, AI Studio, Gemini API, and anti-gravity. Who uses anti-gravity anymore? I don't know. But the price is about to go up November 30th. So you have about 6 7 weeks to really take advantage of the $2 in, $10 out before it goes up to $4 in and $20 out per 1 million tokens. So to put into perspective on what it costs. So per 1 million tokens compared to other Frontier models, you can see Argon the intro is about $10 per output. Sonnet 5.5, which is a substantially faster model, is also $10 out. Argon after November 30th is going to be $20 and Opus 5.5, for example, is $20 out and GPT Astra is $50 on the output

3:03tokens. Still 1 million token context window. So here's what to watch out for. It thinks a lot. 110 million output tokens to run on artificial analysis tests. Median model was 85 million thinking bills as output. The cached input was 95% off. Google's own scorecard here. And this is where I want to talk about the benchmax results. This is insane. If you know anything about benchmarks, you can see here, wow, vows index beating all of the Frontier models. You can see um Deep Sweetbench on Agent Coding beating all of the models. This is amazing. Wow. There's a Vibe coding benchmark beating all of the Frontier models. We're talking about GPT

3:48Astra, Fable 5.1, Opus 5.5. So, if you look at this and you're getting, you know, you're new into this space, you're going to be trying to focus on building with the best model that's the cheapest. And by looking at the benchmark, you're going to believe it, but after using it, you're going to be disappointed. So, Google's own scorecard here, it's a top score on 13 of the 19 evaluations, tied number one on CWEB, and that is pretty interesting. vendor reported Google ran Argon itself. Big mistake. Rivals numbers are on each company's own at max reasoning on a single attempt. Where it wins and these are Google's own reported numbers, not an actual my personal or anyone else's experience is the agentic

4:33coding. You can see Argon scored 77% over Opus, Astra, and Fable on real coding tasks. graph walks which is huge context reasoning over 1 million tokens of data scored 84% which is substantially higher than any other frontier model and then Harvey's legal agent benchmark which is at 19.6% 6%, you know, 10 15% over Fable, Astra, and Opus 5.5. And here is where it loses. And I think that people are starting to understand after testing this model over the last 24 hours, you know, it can be a good model at the price point it's at now, but since they're going to be raising the price in 6 weeks, it's not going to be worth it. A $200 Cloud Code Max plan is going to get you much more usage and better coding, better website

5:18design, and a better overall experience into building your product than something like an Argon where you're going to have to go back and deal with a lot of hallucination. You can see by referring back to Theo's post that he's been testing it all day and is not happy about it, not happy with it. So, where it loses, it's losing on Frontier Sweetbench, Terminal Bench 4.0, and Terminal Bench Science uh.1. So the pattern long terminalheavy coding agents that's where Claude and GPT still lead. So Bloomberg reported some Google staff found it weaker on realworld coding than the benchmark suggest. I'm so tired of these models getting dropped and the benchmarks look amazing but then when you go and use these models they're actually terrible. What independent testers say and these are not Google's

6:04numbers. So for example, artificial analysis the index it says it rated it at 53 which is five points below opus 5.5 three points below sonnet and on the same tier with fable 5.1 and Astra which is still yet hard to believe. So GPT and claude at max effort and then argon on high and then the vows index which is just finance legal and coding and tax it scored 68.9% which is insane. Um, that's why it's just hard to believe these numbers just from benchmarks alone. Top of the pack on knowledge work, middle of the pack on raw intelligence. Cheap to actually run for now because like I said, the price is going to go up. Artificial analysis cost is about $1.99 per task, which if

6:49you're comparing it to Fable 5.1, substantially cheaper. And you can see on the Val's cost per test 15.68 versus like a Opus 32.14. What people are finding Reena.ai AI Gemini 4 Argon High by Google DeepMind just landed number one in the text arena at 1525 points in number eight in code arena. Scow skip on X said Astra dethroned Gemini 4 Argon is the best vision model number one for detection tight precise boxes dense scenes number one for text extraction small text handwritten text complex tables rotated high four times cheaper than Astra on high and then the skeptics Bloomberg it says some of Google's stat found it weaker on real coding tasks like I had said critics say it's benchmarks which

7:35you'll see all over at X right now which is true those are straight up facts And then Google says that's inaccurate. We've been waiting for months for Google to release this model and they released it and it's lackluster. It's something that I wouldn't be using. Here's where I would test it in my resell business. Like if I were to actually use this inside of my main business that we run a live sale business on Tik Tok shop and on whatnot. So they said it was really good at vision, right? So like the pain point for us is reading tiny size and style tags on every piece at intake. So to be able to identify tags that are, you know, a Lou London legging that's a size six, you know, am I able to pass that through a vision model to be able to add it to our inventory index? It says it's claimed in number one at small

8:21and rotated te text and the test would be photo of 50 brand tags and does it get every size and style correct? Another pain point is counting thousands of pieces when a shipment lands. So, it's claimed number one at detection in dense scenes. So, photo of a pile, can it compare that of a pile of a bunch of leggings versus me counting it one by one? That's going to be difficult. Show numbers typed in by hand ever a after every stream. So, strong on charts and tables. Can it build up a dashboard where it can actually, you know, build something out that's going to be usable and readable for me? And I would love a little feedback on are you using this model? What are your thoughts? Please comment down below because I'm trying to

9:06figure out is this model benchmax because I am sick and tired of it. Like I said, I've been testing it just a little bit on the day-to-day. I wanted it to simply just rebuild my landing page. So, my landing page for Shipping School is built with packed with like Opus 4.8 4.9. I, you know, wanted to change the cursor and I wanted this really cool scrolling effect. So, I had that implemented just really simple. Here's the original. And then I had Argon recreate this. This is our free community, by the way. So, I wanted to see how it would do. This is the Argon example here. It looks very much the same, but it didn't do the cool scroll effect. There's this weird blank middle

9:52part of the website that I don't even know that looks horrible. Um, you know, it got this little thing right, but it didn't zoom in. Um, it has this cool little effect that's cut off on the sides. Um, this is all the padding and the spacing is weird. So, all in all, I mean, if this is a Frontier model, it looks terrible. What is this? There's nothing there. And of course, we have to put it through the amazing lava lamp test. This is just a test that I like to do just for fun. And you can see here it actually didn't do very It actually did okay. Um, so as far as like game development and asset creation, I'm still pretty skeptical because Opus 5.5 is so good. But hey, this took like 13 seconds, 15 seconds to create. Um, not

10:39terrible. Um, does it really prove to us anything uh, you know, with the lava lamp test? Nothing really. But at the same time, it didn't produce a terrible result. So all in all, Gemini Argon 4, I think that Google needs to step up their game. I I just feel at this point Google is just playing around. You know, it's not really taking this whole thing seriously. Like it I mean Google is a multi-, you know, trillion dollar company. They don't seem like their heart is in it. And a lot of these benchmarks that come out, even with, you know, OpenAI and Anthropic, they just seem benchmark. So the best real world application for you to do is give it a few tests. Have it refactor your codebase in a separate git work tree.

11:24Have it recreate your landing page to see if it made it any better or made it any worse. There's a few things that I would put it through its paces before I let it go all out on my codebase. Like I said, going back to the start of the video, Theo said it's not good. He says this isn't unusual. Oh, it's 5% worse in this way or 3% better in others. It's almost in every test that he ran, Gemini was factory wrong about over 50% of its claims. This model is a straightup liability. Keep it away from your code. Scary. So before testing these models out, make sure that you're doing it locally and that you're not using any production code to make changes, especially if you have users. Like I said, comment down below. I want to know your thoughts. And if you want to join a

12:10community, a free community of other builders, I'll put that link down below. It's called shipping school. We also have a free discord. We have 171 members. We also have a classroom full of courses, resources, and pass live coaching calls. And we do about three to four boot camp calls a week. So, you could check that out down below. Join a community of builders, testing new models, learning game development, building cool things with AI, and building real businesses. So, what are your thoughts? I don't know. I think I'm going to stick with my Claude $200 a month subscription. It has served me well. We'll see you in the next video.