NOVO Muse Wonder Full da Meta Promete Superar Opus 5.5 e Vira o Rei de BenchMark

Author: Inteligência Mil Grau · 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: NOVO Muse Wonder Full da Meta Promete Superar Opus 5.5 e Vira o Rei de BenchMark · Channel: Inteligência Mil Grau · 24:28 · Watch on YouTube

Topics: AI evaluation and benchmarks, AI content generation, AI models

Summary

The video reviews Meta’s Music Wonder and Music Wonderful models, contrasting their strong benchmark claims with the creator’s hands-on tests. While the models score highly on several reported evaluations, the creator finds their generated websites, illustrations, and games comparatively simple and less capable than outputs from models such as Opus, Sonnet, and GPT 6.1 Sol. The video argues that this gap may reflect benchmark optimization, or “benchmaxing,” and questions whether Music Wonderful is worth its high subscription price. It concludes that Meta’s launch has not yet delivered the frontier-level performance its benchmarks seemed to promise.

Key points

Meta presents Music Wonder and Music Wonderful as models for personal and professional tasks, with Music Wonderful access initially limited to higher-tier Music Code subscribers.
The benchmarks cited in the video place the new models near or above competing systems on several evaluations, including memory and agent-oriented tests.
The creator notes that benchmark results do not necessarily predict performance on unfamiliar, real-world tasks and discusses “benchmaxing” and the “Big Model Smell.”
Some user comparisons on X reportedly show Music Wonder producing simpler results than GPT 6 Astra or other competing models.
In the creator’s pet-shop website test, Music Wonder produces a functional but basic result, while GPT 6.1 Sol generates a more detailed and polished site.
Illustration and game-generation tests show the creator’s proposed distinction between first-, second-, and third-generation outputs, with Music Wonder’s results judged less advanced.
The creator says the models’ practical outputs fall short of Opus and Sonnet, especially in detail, visual quality, animation, and game complexity.
The video argues that Music Wonderful’s reported price—about R$259 per month for the relevant plan—does not seem justified by the creator’s tests.
The creator ends by comparing Meta’s benchmark claims with previous cases where models were promoted before their real-world performance was clear.

Chapters

Start	Chapter	Summary
00:00	Meta announces Music Wonder	The creator introduces Meta’s new models, their ambitious positioning, and the question of whether their benchmark results will translate into real performance.
01:00	Availability and benchmark claims	The video describes access to Music Wonder and Music Wonderful, then reviews the models’ reported scores and rankings.
04:41	Memory, focus, and benchmark caveats	The creator covers Meta’s claims about active memory and focus, and explains concerns about benchmark optimization and performance on unfamiliar tasks.
08:20	Early user comparisons	The video contrasts Artificial Analysis listings with social-media reactions and examples that appear to show simpler results from Music Wonder.
11:30	Testing a pet-shop website and an illustration	The creator tests Music Wonderful on a pet-shop site and a pelican illustration, finding the results functional but less polished than competing models.
14:37	Comparing generated games	The creator compares Music Wonder and Music Wonderful game outputs with other models, describing the Meta results as relatively basic and blocky.
18:43	Generation levels and frontier comparisons	The video uses outputs from Riko, Opus, and Sonnet to illustrate more detailed game generations and argues that Music Wonder falls behind.
20:48	Kart test and final verdict	The creator compares kart games, questions Music Wonderful’s subscription value, and concludes that its practical results do not match its benchmark reputation.

Quotes

Testing the model at home with your own problems is very easy. — 06:15

And they even remind us that there is a concept called bench maxing — 06:15

But for a frontier model, this delivered very little. — 12:32

The discrepancy between the benchmark and what we are seeing is a clear example of benchmaxing. — 21:19

That other one is a joke, it’s just messing around. — 22:22

Concepts and entities

Muse Wonder, Muse Wonder Full, Meta, Opus 5.5, GPT-6 Astra, benchmaxing, memória ativa, foco ativo, Artificial Analysis, Opus 5

Full transcript (original language)

00:00 Yeah, guys, Meta just released Music Wonder. That’s right. A new little model from Meta is out. They also released Music Wonderful, which is an even more advanced version. And that question remains: will they actually make a truly advanced model this time, or will they cement themselves as the kings of benchmarking, where they show absurd scores that turn out to be a huge disappointment when we test them, leaving us wondering how the researchers even had the nerve to

00:30 publish something like that? So let’s understand what is happening. Come with me. That’s it. Let’s go, everyone. Always thanking everyone who liked, everyone who subscribed, and a special thanks to all the channel members who sponsor this artificial intelligence channel. Always remembering that members have access to exclusive videos on intelligent agents where I integrate with WhatsApp, MCP, PDF reading, spreadsheets, and much more. And for those who are apprentice members, there

01:00 is the advanced training. So guys, what’s going on? Music Wonder has been released, which according to Meta is for personal and professional tasks and is heading towards personal super-intelligence. So they are saying here, look, there is the Music Wonder model and also Music Wonderful, which is a play on words there, I don’t know if you caught the Wonder and Wonderful pun. And it is already available both in the Meta API and in the Music Code plan. And Music Wonderful is already available for Music Code subscribers.

01:32 But I need to show you that it’s only on that $ 250 plan and up, at least for now. You won’t be able to use it before , but if you’re on a lower plan, check if it has appeared there, but it has to be on Music Code. The Music Wonder, for those who like to use Mus, will be released first to subscribers, but later on, they will release it to everyone. Look how interesting this journey of Meta’s is here. According to Artificial Analysis, they had 31 points on Music Spark, got 34 on Music Spark

02:04 1.1, scored 40 points on Music Spark 1.2, 48 points on Music Spark 1.3, and now they are scoring 52 points with Music Wonder and 56 points on Music Wonderful. So, technically they are here like Mario Brothers, jumping from brick to brick until they reach the front line between Cloud Fable 5.1 and Cloud OPO 5. So, folks, the only thing we know so far is that they’ve set

02:34 expectations sky-high. Even ChatGPT itself is struggling to position itself here next to Opus 5. And also next to Claude Fable 5.1. And there’s a huge battle here with GPT6 Astra. And in the Gem 4 video here, from what I showed you, Gem 4 isn’t doing as well as they’re saying. The folks at Arena AI tested it and didn’t like it that much, but anyway. Could this be the new king of benchmarks? Hold on, we’re going to test it now. So here now they start

03:06 running a series of frontier tests, and here in the Music Wonderful Max version and also in the Music Wonder Max, which is in blue, they are beating Oppul 5.5 and GPT6 Astra by a certain margin, right? Actually, Mus Wonder is at the same level and Wonderful is winning, look. Practically the same thing here on the Automation Bench, and on the Memory B here, look, they are way out in front, both together here, better than everyone else, including Claude 5. So, the expectation keeps rising, right

03:38 ? What can we expect from these models? The best results, at least in the SI Marathon, are the same thing. It’s ahead of everyone, ahead of Apple 5.5 in deepsui, same thing, ahead of everyone, keeping our expectations sky-high for it to be an absurd model. And here are the same results in table form, right, where they claim to be the best on the planet here, with Mus and Wonderful Max here by a landslide. Where, occasionally, on one benchmark

04:08 or another, right, in this Health Bch Pro here, the winner was Claude Opus, and in the GDP PDF the winner was Astra 6, but overall they are practically keeping up with everyone else at the same level. Either they win or they stay very close. There is no major value here, no huge discrepancy that sets them apart from the others. Although I didn’t see GDP Val here, right? I always talk about GDP Val, because that’s where we know if the model is really good, if it can deliver a complete task there, and it didn’t appear here. But okay, we continue here

04:41 with high expectations. On the frontier capabilities at Harvey Lab. They are also in first place here, look, scoring 13%here, which is already starting to be a much higher number than GPT6 Astra . And they are doing well with both Music Wonderful and Music Wonder. And here in Vals Finance, they are tied with everyone else here, close to 60%. So here, at least, they are more or less equal. They are saying that one of its features is the so-called active memory, which helps to keep important

05:12 context about you throughout conversations and retrieve relevant details. And that is why they say that in the Memory Bench here they are doing very well, better than everyone else, because they focused on this active memory talk. And the other thing they are talking about is the so-called active focus, where Musy Wonder keeps working toward a challenging goal. This here is objective-based tasks, we’re already familiar with it, right? Which would be that thing: it doesn’t lose focus as the task progresses, and so, in the S Marathon, it would be leading ahead of everyone. And in the Focus

05:44 Bench, MUS Wonderful would be leading, while MUS Wonder is there in the middle of the pack, still at the level of the 5.5. Expectations just stay up there. And they insist here that in the deepseek programming part they are also better than the 5.5. Better than the GPT6 Astra, better than everyone else. And well, don’t ask me why, but they included a text here that usually no one includes, but they raise concerns about models that show similar

06:15 benchmark scores, but where one of them seems to handle a wider range of unknown tasks, which is that thing, right? Testing the model at home with your own problems is very easy. I want to see you release this model to the public and those good results you showed there work for the public too. And they even remind us that there is a concept called bench maxing, which is exactly when you optimize your model for that benchmark, get high scores, and in practice your model isn’t that good. And they also talk about another,

06:47 now positive expression, which is the Big Model Smell, which would be something more or less like that, that scent of a good model, you know? which is more or less when the model doesn’t do so well on benchmarks, but in practice it rocks. And so, they invested so much in this analysis to know if their model is good, both internally and externally, that they even showed this graph here, making a comparison with tests within the target , which would be this bottom bar here, which goes from 0 to 100%, and testing with problems outside the target, which is this bar here on the left, which

07:18 shows the performance of the models being used outside of the target training. Technically, the further to the right and the higher up, the better it performs on both internal and external problems. So we can see here that while they’re hitting 70% accuracy on internal benchmarks, they’re at 50%on external ones, which is expected, as you typically have higher accuracy internally than externally. But in a way, what they’re trying to say is that this

07:48 curve, compared to Mus Spark 1.3, has gone up; it has increased, meaning it’s getting more right both internally and externally, which would be very good, after all, that’s what we expect. And another thing they also end up showing is that they are safer. So, for example , while GPT6 Astra has a nearly 20% probability of executing unauthorized code, Mus Wonder and Music Wonderfall are around half that, at 10%, and Cloud Fable 5.1 is also at 11%. And here,

08:20 regarding slips with personal information, it’s the same thing. And it’s interesting to note that when we officially go to Artificial Analysis, the Music Wonderful in the Max version is actually appearing here. It states there is no public API because this model, so far, has only been released to Music Code subscribers. And it is positioned between Cloud Sonet 5. Cloud Fable 5.1. And here with 52 points, just behind Gemini 4, is the Music Wonder Max, ahead of the GPT 6.1 S Max.

08:52 It is interesting to see that Artificial Analysis itself is endorsing these benchmarks of theirs, showing that it’s all accurate. What they showed is appearing here and it all checks out. However, when we go onto X, the story changes a bit. Look how interesting this is. Anchu said: “Music Wonder might outperform Astra in benchmarks, but it is far from reaching Astra in terms of flavor.” Benchmarks don’t tell the whole story. And then he shows here a little animation he made of two clocks. There are two gears

09:24 there that are rotating, that are being simulated. Look at this. Wait, wait, wait. Look here. The one on the right is the GPT6 Astra, doing the internal part of the clock full of gears, running quite nicely, with a lot of detail. The one on the left is the Mous Wonder, which made this thing here—I don’t even know what it is—which clearly has a simpler aesthetic, and even the joints here don’t seem to make much sense. Although you could say that all of them are a bit weird, right? They always have some kind of hallucination there, things that don’t make any sense. But either way, you can

09:56 tell that Wonder has a certain simplicity compared to GPT6 Astra. And here Bav says the same thing. I’m fed up with Mus Wonder. I tried hard to get good results and kept hoping it would deliver something better at least once, but that’s just not happening. Sorry Meta. Based on my tests, it performs worse than Opus 4.8, but costs as much as Opus 5. To be precise, it’s far from reaching Kim 3. And that’s when we realize, right? Kim 3 made a plane

10:27 landing on a runway quite beautifully, while Mus Wonder made a rocket flying sideways landing on an airport lawn. So yeah, Kim 3 put on a show here, didn’t it? Mus Wonder has a lot to learn. And the same thing here compared to Mus Wonder and Kim 3, on the rocket launch, you notice the level of detail in Kimica 3 is much higher and that Mus Wonder is a bit plain, it works, but it’s also very simple. And likewise here with a whale jumping, Kim K3 made a whale and Musy Wonder made a flying

10:59 submarine. It’s a very interesting thing there, huh? That Mus Wonder still has a lot to learn, but all the previous comparisons were always with Mus Wonder and none of them mentioned it was Mus Wonder Full. And this time here using Music Wonder Full, Apple 5.5 made a simulation of a little tractor inside a farm. Look at the quality of this mansion they are showing, right? And look at what we are seeing there in Music Wonderful. Apples made a real game and Wonderful made a little joke

11:30 game there, for us to have fun. Just very simplified, very scaled back. And in our Matrix pet shop test, look how crazy that is. It made a very simple generation, it didn’t generate the images I expected it to. After all, it has an image generator. We can add items to the cart here and this bunny appeared. This bunny, guys, we already saw in the GLM 5.3 or 5.2 tests. And that, to me, is not a very good sign, guys. The website’s image is much

12:01 better, but it would be very strange if someone had used GLM here. I don’t know , because this thing looks like GLM, but we’ll keep adding items to the carts. We can run a test here, see, to trigger the awakening by taking the red pill, check it out. generating a card. The bunny will appear again. This card is super, super simple, it has nothing to do with a frontier model. And when we go to the cart, the cart is also super simple here. We can finish the purchase and it goes through to the end

12:32 as expected. So everything is fine, the site is working. But for a frontier model, this delivered very little. And one of the things I will start doing from now on is showing a reference, showing our top pet shop. so that we can see what is happening. So, check this out. Look at the difference with GPT 6.1 Sol generating a site. In this case here, it already generated the images. The site is much more interesting, it doesn’t have that GLM 5.3 bunny. You can add items to the

13:04 cart here. If you are going to awaken the pet, you put your Bob here. Awaken your pet. There’s even a little game here, see, that you go through to be able to play. And at some point here, it calculates your score and shows you your card with the image all set. Much more interesting, right? What do you think? And when checking out, everything is also very well done. You can fill in the information here, add or remove products, review the order, and everything is purchased and working . Now with our pelican here, the

13:34 drawing turned out quite nice. The pelican’s face and the pelican’s wing are just right. There’s a sun in the background, but for example, the left foot is on the right side, right? The left pedal is on the right side, it isn’t sitting properly on the seat, the wing isn’t holding the handlebar correctly, and the whole scene is very simple, very much like a beginner’s work. The only thing it really got right here is this rear cassette and the rear chain. This is something models usually get wrong, but it got it right. Now compared to our top-tier one

14:07 here, generated by Sonnet 5. Look at this pelican, right? Even though the cassette is facing forward, the left foot is on the left side. It is sitting perfectly on the seat, it is holding the handlebar just right. There’s even a fish, the details, the scenery, everything in the background is much more interesting. And it shows the difference between a non-frontier model and another that is truly a frontier model, which is already getting many details right. And important things, this gear is wrong, it’s technically

14:37 not in the original position, but as a workaround, it works. Now in our little car game, look how interesting, this was the only version I made using Wonder without it being Wonderful. And this one ended up in the max version. And look at what happened. Here is our little car. I’m already calling this generation one of the car game. What is a generation one? Where objects are square. It looks like those 2025 generations. Everything is square. Although this little square train here is very good, isn’t it, guys? Back then

15:08 the little train would drift, but it has no textures on the ground. The game is all geometric, very simple. The AI is totally lost playing there, look. It can’t grab the coin. That is a typical generation one thing. No texture at all . Although the advantages here are that the track is just right, something they used to get wrong back then, and the little train is also just right, despite being quite simple. And so, to be fair, I started making the generations using Wonderful. And here in Wonderful, things changed a bit.

15:38 Although it’s square just like in generation one, now at least there is texture on the ground and also texture here on the asphalt. The little train is cute there, more or less on the same level as the other, right? In this case here, it didn’t change that much. At least the little train isn’t drifting. There are already some buildings there in the landscape and things improved a lot. But the car is still a square car, nothing special. It could have had texture everywhere and it didn’t. It doesn’t look like a frontier model. So, what did I do? I did the magic prompt

16:09 and made some versions and some generations in a version two and then in a version three here making improvements. Did the game improve? It improved. There are more things here, look. The little train now has little smoke. There are some extra details here. But apparently the game remained the same thing. Until I discovered that I was using everything in extra high and that the benchmark models were the max models. And then I discovered that I had to manually write/effort max, and only then was I able to access the Max

16:41 model. And so finally in Wonder Turbo here, already in the max version, it has texture. Things have improved quite a bit, but generally speaking, it’s still that blocky stuff with little blocky trains—classic generation one material. Although the little train isn’t drifting and the track is wonderful, so that’s top-tier, right? That’s the good part, but there’s nothing here that makes me say this is a frontier model from late 2026, you know? The little car is hallucinating its trajectories over there. It’s doing

17:11 it, but you can tell that even this trajectory is quite simple, and although there are more buildings now and things got more interesting, they did. But everything is still very square, all generation one, lamp posts in the middle of the street, those typical things of a model that’s just starting out. And just so we have an idea and a comparison, this here is a version of Sol 6.1. Look at the difference, folks, right? We’re on the Sun here, look. There’s already a little plaza in the middle of the map, and our locomotive. Look how interesting. The locomotive actually

17:41 looks like a locomotive now. And look at our little car. It’s already all rounded; it’s no longer just a little box moving around. It is still based on rectangles, and the rectangles are rounded. So you could say this is generation two. Generation two has this characteristic: texture everywhere. Things are starting to get more or less rounded, and the little games are already starting to look good, right? Yeah, look, the AI got stuck there, trying to escape the lamp post, but since it can’t, it won’t move from there, see? Let’s see if we can help it out here. I’ll get in its way a little more. Maybe it can find a more

18:11 interesting path and get unstuck from the pole. Nope, it didn’t manage to. There was no way. But just so you have an idea, this is another version of generation two. Little cars are also rounded. Look at that nice, little rounded car, perfect. It’s a rounded rectangle. It’s a rounded rectangle. But let’s see how this little train is doing. Look at this little train too. Same thing, see? The locomotive is already pretty, all colorful, and detailed. The track there is also very beautiful, wonderful. Little puff of smoke coming out. This here, folks,

18:43 what model is this? It’s Riko 5. Do you have any idea, guys? A tiny model, Riko , super cheap, producing a generation like this, where the car itself is already fully rounded. So Riko is already doing generation two car games here, while Mus and Wonder are doing first-generation car games, right? It’s a very big difference. But you must be wondering, after all, what is the third generation? I was just waiting for you to ask that. The third generation, look at the difference. We are on Opus 5.

19:16 The little car is already all rounded. It’s no longer just a simple rounded rectangle. The whole car already has a different shape. Look at this track, guys. It already has waves, hills, and dips. There are houses, everything is pretty there, look. The streetlights are not in the middle of the road. This here is a generation with exactly the same prompt. Look at the little grass here, look. Waving wonderful little coins. This is truly a frontier model. Look at the train signaling and the train arriving. Look at the difference of this train arriving here, guys. It

19:46 already has a bunch of cars, all with text already. The level here is much higher, right? Oh, look how interesting this wonderful little train is. Look at these stones. Look at the texture of the stones. When we think of a frontier model from late 2026, it’s something at this level. Look at the little arrow here of the car’s path passing by me. It’s another level. Everything is different here. Everything has a little bit more detail. And the generation of Sonnet 5. The same thing, look. It’s already rounded. Look how interesting. Although this car is a bit ugly, it’s a

20:18 bit weird, but at least it’s also already rounded. And the little train from this version, look. Look at that. Look at this scenery here in the front. I already have a little square here. And the little train from Sonnet 5. It was the first time it made an entire convoy that has the locomotive in the front and has so many cars that the thing barely fits on the screen. So that’s the thing, right, guys? You can’t call this a frontier model. We can’t say this thing is on the same level as Opus, nor Sonnet, nor even Haiku. It’s not worth those 250 dollars

20:48 for anything on this earth. If you want to pay that amount, it wasn’t me who recommended it. Comment down below what you are thinking. Who had the best result? Music Wonderful, Sol 6.1, Opus 5. Sonnet 5. Haiku 5. And in our kart test, look how interesting, the karts are all working perfectly, but you can already tell that the quality is more or less the same, right? Look, it didn’t turn into a great game, it doesn’t have super animation, there are a lot of things that are quite

21:19 simple, although everything is working, right? So, what you can tell is that it is working, but it is far from being a frontier model, it is far from surprising us, since we were expecting something at the Opus level, because that’s what they showed in the benchmarks, right? The discrepancy between the benchmark and what we are seeing is a clear example of benchmaxing. They made something optimized for benchmarks, but it doesn’t deliver in practice. And just look at the difference with an Opus 3.5 . Kart. Look at that difference, folks.

21:51 Here there’s a crowd, the cars work, the visuals are wonderful, there are lots of animation effects there, it’s a hard game to win. I’m accelerating here and no matter what I do, I can’t catch up to them. To actually win this game, I really have to play well. So, it’s safe to say that the graphics are better, the difficulty is higher, the beauty of the game is greater. So, yeah, there’s just no comparison, right? If we compare this game we are seeing here, with the difficulty, the beauty, and put it side by side with the “Mus Wonderful,” this one puts on a

22:22 show in terms of animation and everything else. That other one is a joke, it’s just messing around. Comment down below which one you liked more, this version from Opus 3.5 or the “Music Wonderful” one. I’m biased to say, because this version here is very good. And I didn’t even prompt it or anything, folks. It’s exactly the same prompt, same output. The only thing that changed was the model itself . And one interesting thing that is happening, at least for now, is how it stands. For you to have access to “ Music Wonderful,” you need to be on

22:53 this Power Usage plan of R $ 259 per month. The previous plans here, they don’t include the “Wonderful,” you will only be able to use “Music Wonder. “ And you guys remember, right? This generation of the kart game here is from Mini 2.5 back at the beginning of 2025, right? This one was not missed at all. Who would want to pay to use that? It’s better to just subscribe to the R $ 100 Anthropic plan and use Raiko to make that generation we already saw, right? Even Riko is making more interesting generations. So, apparently

23:25 , it seems like Meta is not in the AI race. It’s just like Google with Gemini , trying to launch something relevant and failing to do so. So, in reality, the king of the benchmark, the number one spot, undisputed, goes to Meta with MusicGen. Because it’s been a long time since a company promised so much, claimed they were in first place, and then delivered a test as messed up as this, while I’m struggling to get the model to produce results. I haven’t been doing any magic prompts for the car game anymore just so I don’t have to keep improving it. And in this case,

23:57 I made three versions and it didn’t help. And in second place here is Gemini 1.5, because they are also in on this, right? They released a benchmark before actually releasing the model. And the people who tested it, who used it—you might have seen it in the video I made—those who tested it didn’t say it was all that great, right ? And in third, who couldn’t be missing is Elon Musk, with his habit of saying he does wonderful things, and for now, Grok is a promise that hasn’t been fulfilled, even though 4.7 was very good, right? At least for generating some games, it was very good, but it’s

24:28 very unstable, so you can’t use it for day-to-day tasks. Go ahead and comment below what you’re thinking. And if you want to support the channel to keep seeing videos like this, become a member. Members have access to exclusive videos on intelligent agents and early access to videos. That’s it, leave that like. Thanks,