Gemini 4 Argon Beats Opus 5.5 But Claude Fable 5.5 Is Coming!

Author: Universe of AI · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Gemini 4 Argon Beats Opus 5.5 But Claude Fable 5.5 Is Coming! · Channel: Universe of AI · 11:25 · Watch on YouTube

Topics: AI evaluation and benchmarks, Software engineering with AI, AI content generation, AI models

Summary

The video covers the global rollout of Google DeepMind’s Gemini 4 Argon and compares its performance with Opus 5.5 across coding and 3D-generation examples. The presenter says results vary by test, but highlights Argon’s realistic-looking outputs and improved coding, while noting that Opus performs better in some cases. The video also discusses Argon’s pricing, availability, and the possibility that output quality could decline after launch. It closes with speculation that Anthropic’s Claude Fable 5.5 may be released soon, based on its appearance in model catalogs.

Key points

Gemini 4 Argon is being rolled out globally and is available through the Gemini API, Gemini app, and Anti-Gravity for eligible subscribers.
The presenter describes Argon’s results against Opus 5.5 as mixed, with performance depending on the test and prompt.
Argon’s introductory pricing is stated as $2 per million input tokens and $10 per million output tokens until November 30; afterward, the video says it will cost $4 and $20, respectively.
The presenter emphasizes Argon’s ability to produce photorealistic-looking outputs, including 3D scenes generated with code.
In a Minecraft-clone test, Argon implemented working features, but the presenter judged Opus 5.5’s result better overall.
A coding comparison showed Argon completing a task in 13 minutes for $1.02, while Opus 5.5 took 15 minutes and cost $8.02; the presenter considered the outputs roughly comparable.
The video suggests that combining Argon with Nano Banana and iterative refinement can produce detailed 3D game-like visuals.
The presenter cautions that Google models’ output quality has sometimes declined over time after release, and hopes Argon will not do so.
Claude Fable 5.5 appearing in model catalogs is presented as a possible sign that an Anthropic release is imminent.

Chapters

Start	Chapter	Summary
00:00	Gemini 4 Argon rollout and early comparisons	The presenter introduces Argon’s global rollout, availability, early comparisons with Opus 5.5, and Google’s plans to accelerate model development.
03:04	Pricing and early visual examples	The video explains Argon’s introductory and later pricing, then showcases a photorealistic office-to-racing-game example.
05:10	Coding and 3D output comparisons	The presenter reviews a Minecraft clone and several 3D tests, comparing Argon’s functionality, realism, speed, and cost with Opus 5.5.
08:19	Detailed scenes and Argon’s strengths	Further examples highlight Argon’s generated 3D visuals and realistic front-end elements, including a comparison with other models.
10:23	Concerns and a possible Fable 5.5 release	The presenter raises the possibility of quality changes after launch and points to Fable 5.5’s appearance in model catalogs as a sign of a potential release.

Quotes

the results have been pretty competitive — 01:01

$2 per million input tokens and $10 per million output tokens. — 03:04

This model took about 19 minutes, used about 6.3 million tokens, and cost about $3.11. — 05:10

I think both of these models are on par — 07:15

this week we are going to get another model from Anthropic most likely, Fable 5.5 — 10:55

Concepts and entities

Gemini 4 Argon, Opus 5.5, Claude Fable 5.5, Google DeepMind, Anthropic, Blender MCP, Nano Banana, GPT-6 Astra

Full transcript (original language)

00:00 The Google team officially just dropped the newest model, Gemini 4 Argon. This model is producing outputs that are better than Opus 5.5. Not only that, it looks like Anthropic is ready to drop a newer version of their Fable model, Fable 5.5. [music] So, let’s get into it. Hey, before we get into the video, I just want to take a moment to tell you guys that AI is not slowing down anytime soon. And if you guys want to stay on top of everything that is happening in AI, be on top of all the new model drops, then make sure

00:30 you guys are subscribed to the channel. Your support means a lot, and now let’s get back into the video. So, today on X we got a post from Logan, basically confirming that the Google DeepMind team is rolling out their Gemini 4 Argon plan globally. Now, this model was announced about 2 weeks ago, but it was kind of being tested at that moment still. They were testing with some cybersecurity defenders, just to make sure I guess that everything is ready to go, as well working with the US government to make sure that their approvals are in place before they launched this model. Now,

01:01 this model is available in the Gemini API for the Pro plus Ultra in the Gemini app and Ultra subs in Anti-Gravity. And a lot of people have been using this model and testing it, and so far the results have been pretty competitive. In some tests, we are seeing that this model is able to outperform Opus 5.5. In some other areas, you see actually Gemini not doing that well or doing a little bit worse than Opus 5.5, but the tests are all over the place. But overall, it’s good to see that Gemini officially has a model that is competing

01:32 with OpenAI and Anthropic. Over this whole year, we’ve been just getting flash models from them. Although those are fun to use and they’re pretty fast and, you know, effective, we still wanted a stronger frontier model from them, and today we actually have that. So, we saw them actually say that we dropped Gemini 3.5, 3.6, 3.7, and 3.8 flash, and our speed of iteration with Gemini is fast, which is kind of funny because it felt like it wasn’t fast. It felt like they were delaying to drop their latest new model cuz Gemini 3.5 Pro got

02:03 delayed, then eventually canceled. And now we are in the Gemini 4 era. Now, what they’re mentioning is that with this new Gemini 4 era, they are planning to increase their speed of development and drop more models. Does that mean the next couple of months we might get more flash models from them and see an update to the flash lineup? Probably. But, we also know that there are some rumors that the Google DeepMind team has solved or I guess are working on RSI now, recursive self-improvement. So, hopefully that does translate to newer,

02:34 stronger models from them because in 2025, the Gemini models were probably the best models out there. And then in 2026, they kind of stopped dropping anything new except for the flash models. And now we actually have a stronger model. Now, the Gemini 4 model, this model is ranking equal to GPT-6 Astra on the artificial analysis intelligence index. And is at a 60% of the cost per task with the discounted price until November 30th. And the pricing for this model currently, they’re doing an introductory pricing,

03:04 which is going to finish at November 30th, which is $2 per million input tokens and $10 per million output tokens. And if you just want to look at the pricing after, if you scroll all the way down on their blog post, at the bottom they actually mention that after November 30th, the pricing is going to be $4 per 1 million input tokens and $20 per million output tokens. And this is going to apply to the Argon model intended for public access only. Now, if you compare this to the Opus 5.5 pricing, it is $4 per million input

03:35 tokens. So, four versus four, which is about the same. And then output 20 and 20. So, you can compare these models side by side because you’re going to be after November 30th be paying about the same for both of these models. So, that is why many of the tests that I’m going to show you are going to compare Gemini 4 Argon to Opus 5.5. And it does make sense to compare to Opus 5.5 at the moment because Opus 5.5 is probably the strongest model available right now. But before I do that, this is an output that kind of shows you what to expect

04:07 from the Gemini 4 Argon model. What we’re seeing is that this model is able to create outputs that are more photorealistic. Many of the outputs I’m going to show you, you can see that if you compare them to Opus 5.5, the outputs from Gemini 4 look more realistic in comparison. And even this output, you can see that this person basically took a picture of their home office desk and then turned it into a playable racing game using Blender MCP and 3.js. And they actually put instructions below. So, if you guys want to figure out how to do this, this is actually a developer person from Google

04:39 DeepMind themself or Google AI Studio. And this is like the whole video about how they made it step-by-step. If you want to, I guess, check this out, make sure you guys go to her account and check it out. But I think this output is pretty competitive for what we have seen from the Gemini team. Now, let’s look at it when it’s being compared by regular users, not somebody who actually works at Google DeepMind. Before we continue, we just launched the Universe of AI newsletter. If you want to stay on top of AI news without having to hunt for it, link is in the description. Don’t miss out. So, here is a Minecraft clone

05:10 made by the Gemini 4 Argon model. And one thing to see is that this model did a pretty solid job of making sure that all the features were working. This model took about 19 minutes, used about 6.3 million tokens, and cost about $3.11. And obviously, this does not look completely like the Minecraft clones that we have seen. I’ve seen better Minecraft clones from probably Opus 5.5 model than this. But the usability and the features and elements are there. So, we have to give some credit to the Gemini model because it was able to add

05:42 all of that. But when it comes to the actual output, if I’m just comparing this test to the Opus 5.5 results that we have seen, Opus 5.5 definitely excels at this test. But what this clearly tells you is that this model has improved on its coding because it was able to create all the features and they’re playable and they work. And when it comes to the output quality, it is pretty much on par with Opus 5.5. But however, on this specific test, I would give Opus 5.5 a better score. Now, as I said, in some areas this model is doing a really great job at creating

06:14 realistic outputs. What you see on screen right now is a 3.js test and everything here that you see on screen is actual code. This was not, you know, using any other external assets to generate this. And when I looked at this output, at first I thought it was like a video generation or something like that, but then it was surprising to see that all of this is code. So once again, the Gemini team has really trained this model to create photorealistic outputs because that’s the common trend that I’m noticing across the, you know, all the

06:44 demos that we are seeing. So now, this is a test between Opus 5.5 and Gemini 4 Argon and both of these models had the same prompt and both of these were set at the highest reasoning available. The Opus model took about 15 minutes and cost $8.02 using the Cloud Code CLI and Argon took 13 minutes and cost $1.02. Obviously, this is because it’s probably using the introductory pricing for the Gemini 4 Argon model, but it’s time-wise, it looks like Argon was able to do it 2 minutes less, which I guess

07:15 is a good thing to see. But if you look at this output, it’s pretty hard to tell which one wins directly from this. As I said, in most areas these models look like they’re on par. In some tests, depending on the prompt, the Gemini 4 Argon model is producing stronger outputs compared to Opus 5.5. But if you just take a look at this, I think both of these models are on par and I would say this one is like a like a coin toss. I don’t know. It’s up to you guys. What do you think? Which one wins in this example, Gemini 4 Argon or Opus 5.5? Here’s another test. Once

07:48 again, Argon against Opus 5.5. And here is a similar story. They are pretty much on par, but I do like the elements that Gemini 4 Argon added. If we look at the exhaust behind the rocket, it looks more realistic compared to the Opus 5.5 one. However, in this specific test, it this was Opus 5.5 set on medium. So, it that could be the reason why it’s not creating the same to level of details like Gemini 4 Argon. But, in this specific test, pretty much a tie, but I’m leaning towards Gemini 4 Argon because the model is producing outputs

08:19 that look a little bit more real or detailed compared to Opus 5.5. Now, this test once again tells you how good this model has gotten at generating outputs that don’t look like simple voxel boxes anymore. This is a futuristic 3D batter helicopter made purely in 3.js. So, everything you see over here was generated from scratch, and this output that you’re seeing is a result of multi-round self-iteration loop of Argon plus Nano Banana. You are combining Nano Banana’s image understanding with

08:51 obviously the stronger Argon model, and by doing that, you’re able to create something that looks pretty solid. This looks like something that you would expect in a, you know, pretty decent game. This was something that you wouldn’t expect the model, especially the Gemini model, to be able to create before Gemini 4 Argon was released. Now, this test is kind of important as well because this kind of puts everything into perspective. Argon was compared against Opus 5.5, GPT-6 Astra, and Kimiko 3. And four of these models were told to build a 3D wheel breaching

09:22 scene. One thing you can see the level of detail that Gemini 4 has compared to any of the other outputs that we have seen, it’s evident that the model is obviously adding more elements, adding more realistic elements. Because if we look at GPT-6 Astra, you can still tell like this is like, you know, cartoonish. Same thing with Gemini 4 Argon looks a little bit more real. Opus 5.5 does a solid job. In this test, I think you can give it to either Gemini 4 Argon or Opus 5.5, but it’s pretty clear that GPT-6 Astra and obviously Gemini 4

09:52 Argon can’t compete with both of these models at top. On the front-end side, it looks like Gemini 4 Argon has a little bit of advantage because as I mentioned and as we have seen many of the times Gemini 4 is creating more realistic elements and that is obviously going to elevate your front-end outputs. If you look at the top and the bottom, you can see Gemini 4 Argon all the elements look way more realistic compared to Opus 5.5. So, I think that is one of the things I love about this model is because of the realism. It makes all of the outputs that it generates feel more high quality. Now, the only caveat with the

10:23 Gemini 4 Argon model or just the Google DeepMind model is on release day most of the times we do expect strong outputs from them, but over time we do see them kind of deteriorate in quality. So, we’re just hoping that that does not happen with this new model Gemini 4 Argon and the level of output that we are getting Google obviously doesn’t nerf it. Now, if you thought Anthropic was going to let Google DeepMind take away the spotlight from them, you are wrong because it looks like Fable 5.5 could be coming anytime soon now because it’s actually appearing in the model

10:55 catalogs and this is usually a sign that a release is somewhat imminent because if we look at this, we can see Claude Fable 5.5 appearing over here and it’s not surprising we’ve seen Fable 5.1 last week being routed to a newer model, which is Fable 5.5 producing outputs that are probably going to blow all of us away when this model releases. So, it looks like this week we are going to get another model from Anthropic most likely, Fable 5.5 and when that model drops, I’ll make sure to make a video on it. But that’s it for today’s video.

11:25 Make sure you guys are subscribed to the channel. Follow our new newsletter as well at universeofai.beehive.com as well as to the main channel World of AI, and support us on X by following the universe of AIZ as well. Until then, I’ll see you guys in the next video.