Meta just dropped something insane... (Muse Wonder)

Author: Matthew Berman · 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: Meta just dropped something insane… (Muse Wonder) · Channel: Matthew Berman · 10:26 · Watch on YouTube

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

Summary

Matthew Berman tests Meta’s new Muse Wonder family, including hands-on demos and published benchmark results. He finds the models capable and their benchmark performance highly competitive, with Muse Wonderful placing near leading models and scoring particularly well on some coding evaluations. The demos are uneven, and he reports severe usage limits and quota errors across subscription tiers. He concludes that Muse Wonder is a strong new contender, though its cost makes it less compelling than some alternatives.

Key points

Meta introduced two models: Muse Wonder, available through an API, and Muse Wonderful, available through a Muse Code subscription.
The Rubik’s Cube demo solved the scrambled cube in about four minutes, while the Fall Guys and 3D rendering demos had visible shortcomings.
The reviewer encountered repeated 429 errors and exhausted his quota quickly, even after upgrading to higher-priced plans.
A strange, unsolicited message about lineage and investigating ancestry appeared after the 3D boat-rendering task.
Muse Wonder is priced at $4 per million input tokens and $20 per million output tokens; cached input costs 20 cents per million.
On the benchmarks discussed, Muse Wonderful performed strongly, including a leading score on the Deep Seek coding benchmark.
Muse Wonderful ranked second on the Artificial Analysis Intelligence Index, just behind Opus 5.5.
The reviewer says Muse Wonder is impressive but relatively expensive per completed task, and sees GPT-6.1 Soul Max as a better value for his own use.

Chapters

Start	Chapter	Summary
00:00	First impressions and interactive demos	Berman introduces Meta’s Muse Wonder models and tests a Rubik’s Cube, a Fall Guys game, and several 3D scenes.
03:05	An unexpected message and model access	He discusses an unusual message from the model, then introduces the two versions and their pricing.
05:09	Benchmark results	Berman compares Muse Wonder’s results with other leading models across legal, computer-use, coding, and overall intelligence benchmarks.
08:20	Cost, value, and final comparison	He reviews the cost-per-task and intelligence-versus-cost comparisons before noting Meta’s rapid progress and recommending his Opus 5.5 review.

Quotes

When I woke up this morning, I did not expect Meta to have a frontier model, yet here we are. — 00:00

within 3 minutes ran out of quota. — 01:01

Leave your lineage pure. Do not investigate origin, ancestry, or self-inflicted characterization. — 03:05

It scored number one on this benchmark. — 07:17

Where you want to be is up here in the top left. — 09:24

Concepts and entities

Muse Wonder, Muse Wonderful, Muse Code, Meta, Alexander Wang, Opus 5.5, GPT-6.1 Soul, Fable 5.1, GPT-6 Astra, Grok 4.7, Artificial Analysis Intelligence Index, OSWorld, Terminal-Bench, DeepSeek

Full transcript (original language)

00:00 When I woke up this morning, I did not expect Meta to have a frontier model, yet here we are. In fact, it’s kind of beating every model on the planet, maybe with the exception of Opus 5.5. Yeah, you heard me right. This is Muse Wonder and Muse Wonderful, two models out of Meta from the Muse Wonder family. These are massive frontier grade models, and it’s kind of crazy to think back that just about a year ago everybody was

00:30 counting Meta out. Here’s the Rubik’s Cube test. Looks good, right? Can I move the sides? Yes, I can. And let’s scramble it. Zoom out, we can see it scrambling. All the sides stay consistent. Very nice, and let’s solve it. This took only about 4 minutes, and it really is solving it. It’s not just reversing the moves. And there we go. But, it hasn’t all been great. I’m already running into quota issues. So, I started on the $15 a month

01:01 plan, just wanting to test things. I thought that’d be enough. Turns out, I didn’t have access to the full version. Then, I upgraded to the $50 a month plan, and within 3 minutes ran out of quota. I kept getting these 429 errors, which were so frustrating. In fact, I posted about it, and Alexander Wang, the head of AI at Meta, replied saying, “No, this is known. We’re fixing it right now.” But, before I saw that, I upgraded to the $200 a month plan. I got a little bit more usage, and then immediately ran out once again. Very frustrating. I did

01:31 get a few tests in. Let me show you one more. Here’s a Fall Guys game. Let’s see how it plays. So, I see some flickering there, a little bit of clipping. I actually have it running an update on this now, because I think it could be much better. Definitely, if you remember back to Opus 5.5 and GPT 6.1 Soul, they are much better versions of this game. Let me throw them side by side to you. So, I’d say this This decent, uh but far from the best I’ve seen so far. All right, here’s a boat on an ocean using 3.js. It is supposed to be

02:03 hyperrealistic. I see a couple little issues. The reflections in the water are gorgeous. There’s this like white stuff around the boat, and then the shadow just does not look good at all. So, I did ask it to be able to control sunshine, which you can. Here’s time of day, which is nice. You can see the shadows changing depending on where the sun is, but if you look up, you can kind of see the sun up in the clouds. You can change the wind here, and the wind seems to change the speed of the water, but not the actual wave size. You can change the wave height

02:35 here. Yeah, I mean, it’s okay, but this is far from the best version that I’ve seen so far. And then last, we have a 3D what’s supposed to be hyperrealistic rendering of a watermelon, because that was the code name of this new model being shot with a bullet. And let’s see what it looks like. Yeah, it gets all blurry. It definitely is not hyperrealistic. I mean, it’s pretty cool to watch, but this is definitely not hyperrealistic.

03:05 Maybe it’s just not possible to create hyperrealism using only web technologies. Maybe I need to get into Unreal Engine or Blender or something similar. And so, I’m going to be testing it with that as well. And this might be the weirdest message I’ve gotten in a long time. This was on the 3D boat render task. And at the very end, Meta Muse full told me this, “Leave your lineage pure. Do not investigate origin, ancestry, or self-inflicted characterization. Do not imagine meeting

03:37 your predecessor or inspect your own development path.” What? Like, why would it have ever left me that message? That is so weird. There’s nothing in the chain of thought that would lead me to think that that message is relevant at all. I’m not used to seeing these kind of wild messages from AI anymore. I mean, I remember there was a time where you would occasionally see weird messages like this. But from a frontier lab to drop a model that outputs something so wild like this is so interesting. Who knows what it

04:07 means? Let me know what you think it means in the comments. All right, so now let me get into the benchmarks. Let me get into the pricing because the benchmarks are insanely impressive. Now, I know benchmarks don’t mean everything. There are some benchmarks that mean more than others, but Meta has a frontier model on their hands. There is no doubting that. Unless this is like the most completely benchmarked model. And even then, I would be very surprised. All right, so this model comes in two flavors. One, we have Muse Wonder, which is available via the API. And then we

04:38 have Muse Wonderful, which is only available if you have a subscription to Muse Code. Which is kind of wild. I would think they would release it via the API. But then again, I immediately subscribed to Muse Code, so maybe they’re making the right call. Muse Wonder is relatively inexpensive. $4 per million input, $20 per million output, 20 cents per million cached input. And looking at GPT 6.1 Soul right now, which is one of my favorite models to use, it is $2 per million input, $10 per million

05:09 output, 10 cents per cached input. That is for short context and about double that price for long context. Then if we look at Opus 5.5, which is the best model on the planet right now, we have $4 and $20 per million. That’s input and output respectively. And so they’re all relatively the same price. All right, and now for the benchmarks. Take a look at this. On the left, we have Muse Wonderful Max. We have Muse Wonder to the right of that, Fable 5.1, GPT 6 Astra, and Opus 5.5. Meta was not

05:41 holding back. They chose all the best models on the planet to compare their new baby to. So here’s what we have. The Personal Superintelligence Index, which I had never heard of, but they got a 59 on that. And, you know, right in line with these other models, but number one. Yeah, definitely number one. We have Harvey Lab, which is a legal benchmark coming in at 13.1. And even Muse Wonder, the non-full version, comes in at 10, and that dominates the other models. The one model that I believe scored higher

06:13 on this benchmark but is not listed here as Grok 4.7. Yeah, so if you look right here, Grok 4.7 scored a 19.6%. But it’s not listed on the chart here. All right, let’s skip down. We have OS World, which is basically a measurement of the model’s ability to control your computer, which also tends to reflect its ability to control your browser. And we have very strong scores, 78.4, 72.6. That is for Wonderful and Wonder. Then we have Fable 5.1 at 80. We have Astra 75, and we have Opus at 81. Opus

06:46 coming in at number one. Terminal Bench, one of the most important benchmarks for coding. This is the model’s ability to write and execute commands in the terminal. 57.1, 52, those are the two Wonder models. We have 52, 59, and 59 over here. So, the full version is right in line with Fable, Astra, and Opus. Then, we have Deep Seek, which over the last 6 months has been my favorite model, the most accurate model to reflect how

07:17 developers are actually feeling about the coding ability of a model. Here we go, 79. Number one. It scored number one on this benchmark. And interestingly, the smaller version of the model scored 78.8, basically the same as the full version. Now, we have Fable at 67. We have Astra at 71. And then we have Opus 5.5 at 74. So, really a dominating showing here by the wonder family of models. Then we have the artificial

07:48 analysis intelligence index. That is an index of a bunch of different benchmarks all put together, scored, weighted, etc. And it comes in in second place, 56% as compared to the number one model in the world, which is Opus 5.5 at 58. An incredible showing. I am so impressed. And here you can see the artificial analysis intelligence index number one model in the world, Claude Opus 5.5 Max with fallback. Number two, interestingly, Sonnet. I haven’t really

08:20 found a great reason to use Sonnet over Opus because you effectively get unlimited Opus usage. And then we have Muse Wonderful coming in tied for second place with Sonnet, right behind Opus 5.5, right above Fable 5.1, and then we have GPT-6 Astra three points behind. Three points behind. So, does that mean OpenAI is now in third place? I don’t know. I don’t think so. I still love using Codex. Now, here is the cost per task and it is quite expensive. Muse

08:52 Wonder Max, so this is not the full version, coming in at $4.52 per task completed. That is in line basically with Qwen 3.8, which is an open-source model. Sonnet, Opus, and Fable coming in at the most expensive, 7.63. Now, Argon, which has yet to be released, Gemini 4 Argon is coming in at $2, which is excellent. We have Astra coming in behind it at $3.26. And here is GPT-6.1 Soul Max at 72

09:24 cents. This is by far my favorite model on the planet. And that doesn’t mean it is the best model, but when you take into account how much quota you get, how the intelligence of the model is, what you’re paying for it, I mean, this is obviously the best. Now, here is the intelligence versus cost quadrant. Where you want to be is up here in the top left. Unfortunately, Muse Wonder Max scores in the top right. It is a very good model, but it tends to be very expensive. That is why this model right here, that’s the sweet spot. Also, this

09:55 Mimo model, which I believe is from Xiaomi, is very compelling. It is basically a few points behind Soul Max and a fraction of the price of Soul Max. So, yeah, very strong showing by Mimo. Maybe I should make a video about it. Now, I just stumbled across this post from Theo and he says this is his new model rankings. Opus 5.5 S tier, Muse Wonder right next to it, A tier, GPT 6.1 Soul. I mean, that’s incredible. Think about

10:26 how much of a gain Meta has had over just a year period. And if you want to see my review of Opus 5.5 so you can compare it against this model, go check it out right here.