Meta Just Dropped a Massive Update (Muse Wonder)

0:00Oh my god. Muse Wonder just made the best Subway Surfers game I've ever seen. This is crazy. Wow. That lava lamp is legit best lava lamp that I've seen in a while. Nice. It just recreated my entire landing page in 15 seconds. This is beautiful. Well, Meta just dropped Muse Wonder. And in this video, we're going to be breaking it down on its use cases, what I've built with it, and if it's even worth using. So, Muse Wonder is the newest Frontier model released by Meta. It is a series combines for frontier coding, longunning agents, and advanced computer use to tackle your most demanding work. And like I said, we're going to be breaking it all down in this video as far as like use cases, what it makes sense to use it for, and where I

0:46think it could be better when it compares to other models such as Opus and GPT Astra. So, you need to ask yourself, what is the use case that you're looking to use? Are you looking to build games? Are you looking it for code and PRs and looking to make be refactor your codebase? This is something you need to be wary of to make sure that you're going to be confident in using these new models on production code bases cuz they could essentially screw everything up. So that's why you need to run through a few tests, which is why I do these videos every time a new model drops. So Muse Wonder, it just released 3 days ago. You've seen mixed reviews on X and on YouTube, but I'm going to be breaking it down as maybe why I would be using it versus maybe Opus 5.5 and where I would steer clear

1:31from. So, two models they released. They have Wonder Wonder full and then they also have a Wonder Open releasing in the next few days, which is their open- source model. I do like the 1 million token context window. It's $4 in, $20 out, so pretty comparable to like an Opus. And you can see here two models and one catch. So it's the same name but they come with very different access. So meta model API plus moo code which is muse wonder. And on open router 2 this is $4 in $20 out and then you get the 1 million token context window. So what developers can build on that's something that most of the people watching this channel are going to be using. Muse wonderful is a muse code subscription only. This is not available in the API.

2:16It's $50 a month or $200 a month on maximum plans. There's no per per token pricing. Both are 1 million token contacts. So you get text, image, video, audio in and no open weights. So what is new about this? Meta hasn't been in the whole frontier model race just in this last month and they're quickly catching up. So getting into some of the benchmarks, I do like to go over them, but they're not indicative actual results. So, a lot of these models that drop are super benchmarks, benchmaxed, and they are have their own bias for the benchmarks that they are producing. You could see my Gemini Argon video that I released the other day. It scores amazing on benchmarks, but as far as performance, it's not good at all. So, what's new, what Meta did, and how they

3:03built around it is active memory. So, it remembers what matters about you across chats, which is nice. And then active focus, which is a cool one. It keeps working past blockers and then it tries new approaches and then it checks in before big actions. So on memory bench here, you could see that Wonder scored more than Opus 5.5 and Astra and Fable 5.1. Take that for what you will, but it has the best memory score of any model in Meta's table. The API version even edges out full, which is pretty awesome. So Meta's own scorecard, 13 results versus Fable 5.1, GPT6, Astron, and Opus 5.5. Wonder Fool topped on seven of those benchmarks out of 13 and Wonder through the API. It scored a top four out of the 13. This is all vendor

3:49reported and all of that at Max Reasoning. Some rival scores on Meta's own best effort runs and then others are what each company published. Let me zoom in here so you can take a look. You can see her on Terminal Bench 4. It outscored all the other models except for GPT6 Astra on max and Opus 5.5 on max, which is pretty good. And just to give you some context, I use Opus 5.5 every single day as my daily driver and nothing has come close to producing the results within writing code, building landing pages, design, responsiveness, hallucinations. Opus 5.5 is just amazing. But I'm comparing, you know, a lot of these models to that experience that I've had with using Opus 5.5. And I think Meta, you know, cooked here. I

4:34don't think it's a replacement, but I do think it does warrant some use cases. So where it wins here according to the benchmarks meta's numbers on long coding sessions and legal work. So on the deep SWE bench on 1.1 long horizon coding it scored over opus 5.5 and Astra and then the SU marathon ultra long coding scored about 6.7 points higher than Opus 5.5 and then the Harvey Lab double A legal agent work it outscored all of the Frontier models. Comment down below are you using Moves Wonder? What have you built with it? Do you think it's just hyped up? I think we're at a point now where all these models dropped. They're just super benchmacked. They need to be reviewed and studied to see if they're even usable. So, the biggest lead here

5:19was the marathon on the long coding jobs that run for hours. Now, I want to talk about where it loses in these benchmarks. So, these are Meta's own numbers that they just came out. So, the terminal bench, it wasn't exactly on par with Opus 5.5. You can see here 59.6% 6% versus Wonders 52%. Opus 5.5 wins at all three. Building from scratch is the widest gap. Also, Program Bench built from scratch, not an amazing score. It scored 50% compared to Opus' 70.5. OSWorld 2.1 using a computer also scored almost 10 points under Opus 5.5. So what independent testers saying you know artificial analysis these aren't meta's numbers these are independent labs that

6:04did the testing and on the intelligence index uh you know not bad not bad for meta's first frontier model 52 compared to opus is 58 and you can see just below argon but I had tested argon all day yesterday and I think wonder is even better so it's really hard to take these results on the intelligence index and take it with a grain of salt right so speed is where I think it wins the It scored 161 versus Opus 96. Opus is known for being slow, but I'd rather it be slow and correct and hallucinate less than be fast and then go back and fix errors all the time. So, it's one of the fastest top models. Now, what it costs to finish a task. This is important that are using people through the API. So, cost per task according to artificial analysis. You can see here Wonder is

6:50about $4.52 versus Argon and Astra in obviously Fable or Opus is going to be much higher. So, it's a middle of the pack. It's cheaper than Claude, pricier than Astra and Argon, but Wonderful has no API price at all if you do plan on going through the subscription. You can see here on the artificial uh analysis that Argon is the cheapest model to run, but it's not always producing the best results. So, if you're getting into this space and you're trying to figure out what model do I use, what subscription do I go with, I think that you need to stop messing around and just use what's going to give you the most bang for your buck. Either a OpenAI subscription, which I would probably hold off on if you have the money to spend $200 a month or even $100 a month, get a Claude

7:35Anthropic subscription. Fable 5.5 is going to be coming out in these next few days and you do not want to miss it. If Opus 5.5 is already the industry standard and leader in this whole AI space, can you imagine what you can get done with Fable 5.5, grab an anthropic subscription, but maybe, hey, you don't have $100 to spend. Then you can start looking at like maybe a Muse Wonder or a cheaper OpenAI subscription and start playing around. You can also go through Open Router. A lot of these models are cheap to use like Muse Wonder and maybe put $10 $20 in and start your project that way. I would definitely stay away from any kind of, you know, Google AI uh studio subscription or anything like that. It's just not thorough enough when you start building products. Maybe for design it's okay, but for actual code

8:22work, it hasn't been good in my experience. So, what are people saying? Like the 3D demo drama, people have been building games, a lot of games lately, which is cool. I love to see it. But they straight out came out and said, Alexander Wang says the whole team feels dumb now. But it's true. We did not explicitly train our model on 3D demo games. Obviously, in retrospect, we probably should have done a bit. So, the model is only good as good as what it was trained on. It was not specifically made to create games. Therefore, it's not going to be as good as other models that have some training data. You can see Chris at Chris GPT on the 12 said, "Oopus 5.5 is absolutely mogging Wonder on every single 3D game demo I've seen." Charlie Holt said, "Wonder is really good. found a really a real security vulnerability in our codebase in one

9:08shot that other agents had missed. And then you can see Paul Hearn said tested Muse Wonder full on real work. It caught 105 bugs. Uh GBT6 Astra on Max caught 45 of those bugs. Muse Wonderful Full on Max caught 44 bugs. Fable 5.1 Max caught 43 bugs. And Opus 5.5 caught 41.7 bugs. So it's weak at flashy 3D demos, but it's strong on real code work. So, which is it? Right. So, my test that I specifically did, you know, not anything amazing, but I recreated the landing page, I did the lava lamp test, and then I created the subway surfers game, um, which my kids played for at least an hour this morning, which is remarkable for that cheap of an input. But Meta says, you know, it wasn't trained for any type of 3D game development. So, I

9:54just gave it the same runner through Open Router. I just gave it a oneshot prompt, same prompt, high effort, 48,000 tokens on max. And you can see Muse Wonder um ran for 2 minutes and 43 seconds, 12,000 output tokens. It cost 26 cents to make that lava lamp. While although it's not amazing, it's still very cheap in 2 minutes versus something like a Gemini Argon, which we did the other day, took nearly 8 minutes, 39,000 output tokens, and was 40. I mean, which one looks better to you? And then you can see here it redid my landing page, which I thought it did a fantastic job compared to Argon. So simple. I know Argon had some weird spacing issues like this, like weird padding. It didn't get the concept right. It also didn't look

10:39as clear as it does here. And then it absolutely k just crushed this Subway Surfers game. I mean, although it's not amazing graphically, it still did this in under 5 minutes. Okay, you can see here 4 minutes 28 23,000 output tokens 48. This is something that was just spun up in just minutes. I mean, come on. Is this going to change the world? No. But at the same time, these models are getting better at game development. And I think game development is going to absolutely explode over the next few months because now you can start building personalized games. You can start building games as a community. I think this is going to change the barrier to entry for everybody, which I am super pumped up about. So, do I think you should be using Muse Wonder? I think

11:24that in certain use cases, yes. But if you are already getting into this space and you're looking for the best bang for your buck and you're tired of switching from model to model, just go out, get an anthropic subscription, $100 a month, $200 a month, and you will never look back. I certainly haven't. And I've been managing my whatnot live streaming business with this model every single day. You know, buyer relations, things like that. And if you need help getting started, wherever it is, maybe you're wanting to build a product, maybe you're learning how to incorporate AI with your current business already, you could check out link down below. We have a free community of over 173 members where you can come in, hang out. We also do a couple live boot camp calls every single day. We also have a classroom filled with live replays, builder resources,

12:10and you can build apps through these courses as well. We have from beginner basics to getting cloud code set up to getting started with codecs. Hermes local LLM and AI foundations. Come check it out. I'll put the link down below. You can also join our free Discord. Come connect with me. Come answer ask questions. And we're here to grow the community and to be able to help each other grow in any way possible. Have a blessed rest of the day and we'll see you in the next one. Bye-bye.