MetaのMuse Wonderを検証：エージェント業務は強いが、デザインは苦手

Welcome back to HowEye.AI. I'm Claire Vaux, product leader and AI obsessive here on a mission to help you build better with these new tools. Today I'm reviewing Meta's new frontier intelligence model, Muse Wonder, which comes in two versions, Wonder and Wonderful. I am going to put this model through the HowEye.AI benchmark to see if it can enter the race for frontier intelligence, especially around coding, knowledge work, and agentic workflows. Let's get to it. This episode is brought to you by DX. In a recent study across more than 500 engineering organizations, DX found that spend on AI tools has grown 28x over the last year. The share of AI authored code is climbing, but overall innovation has remained flat. As teams generate code faster, new friction in code review and validation is offsetting those early velocity gains. DX tracks speed, quality, and cost together across the software development lifecycle, giving engineering leaders clear visibility into how AI impacts delivery and whether those

investments are translating into real value. Download the full report at getdx.com slash howiai. That's g-e-t-d-x dot com slash howiai. Let's talk about what Muse is and then we'll get into how it performed on the Hawaii AI bench. So as Zuck announced on X, and I guess probably threads, Muse Wonder is their new family of frontier models with state-of-the-art intelligence and performance on personal and professional agentic work. And if you look at the benchmarks and how they're comparing themselves, they're really putting it in the class of Fable, GPT-6 Astra, and Claude Opus 5-5. Both of those are now my daily drivers. Now, if you look at this benchmark, where they're really grading themselves is agentic personal and professional work. They're saying that this is the best model for driving a personal agent. Now, this makes a lot of sense given their push into Muse, which in my opinion is a very great personal AI agent,

and the model behind it has been zippy and cheerful to be with. So I'm not surprised they're looking at it. But they're not just looking at personal agents. They are looking at professional knowledge work as well. So you're seeing a lot of things around specific domain expertise, like legal work, financial work, business workflows, professional work, etc. So you can see they put a lot of emphasis in their benchmark table on those personal and domain expert benches. But they also care about coding. So they are saying that they outperformed every other model on their bench, including Opus 5-5, Fable, and Astra on DeepSuite and SweetMarathon. So they are really putting themselves in the big leagues when it comes to coding. We are going to see if that holds true and if I would really move some of my agentic coding workflows on to Muse Wonder. But in addition to the benchmarks, they're calling out three capabilities, what they're calling frontier capabilities, that make Muse Wonder a differentiated set of models in the market.

The first is domain expertise. And so they are looking at applying specialized knowledge to problems that people encounter at work. Again, this is where you see them really aiming to perform on legal analysis, finance benchmarks. So they are saying we have specialized expertise in knowledge work beyond coding that we do really well in. The other two are these ideas of what they call active memory. And so they are saying that they outperform all these other models on use and creation of memory. This is very important in personal agents. So I'm not surprised to see that they have come up with something kind of fancy here. But they are trying to carry important context about you across conversations and make sure that across conversations you're getting the right context for you in the task. The other thing that they're really focused on is sustained effort. And they call this active focus. And so when Muse Wonder goes into active focus, that means that it's working towards a challenging goal.

It's trying different things. And so really surprising, I will give you a little spoiler alert on the bench. We did see active focus actually come into play and outperform on some of our long running tasks in the benchmark. So our bench is showing that this active focus investment on our longer running tasks is actually outperforming some of the other models. So that might be a place that I double down later and go into more deeply. Okay, now let's talk about the most annoying part of Muse Wonder. It is that you have to download Muse Code. I mean, I guess you don't technically have to download Muse Code, but this did force me to subscribe to the Muse subscription, get access to the meta model APIs, and download Muse Code, which I will say, in my very light experience, was deeply unimpressive. but did the trick, and I didn't actually have to pay attention to it. So if you want to know how I actually tested this, I did sign up for MuseCode, which you can sign up for at developer.meta.com.

I copied this install script in Terminal, which I have not opened Terminal, I don't know, in three months. So that was a real vibe. And then I gave myself access to MuseCode, and then I had CloudCode run it headlessly. So I did not want to have anything to do. With MuseCode, it's an agentic coding harness. I'm sure it's fine. I don't want to pay attention to it. I just love my desktop apps at Codex and CloudCode. So I lived there. But if you want to try MuseCode, that is completely available to you. And it looks like this. It's not that thrilling. I legitimately have not opened a terminal app in a very very very long time So just know that I did not really test MuseCode I I signed up I gave Cloud Code access to Muse code and ran it headlessly But let's get into how the model performed. And just a reminder on how AI Bench works, we run against almost 20 different tasks.

That's everything from writing a PRD to doing a set of prototypes to redesigning the chat PRD site to running long-running computer use, agentic research, code repair tasks, and then some creative tasks. And then a judge, a LLM of the judge, judge them, and then I judge them. For sake of speed, I'm just going to, like, give you all my hot take feedback on me. So this is really like a first look kind of review. This isn't a benchmark review because I just don't want to run. We just did Opus 5.5 and Sol 6.1 and the whole thing. So I don't want to do the whole thing again. And I just want to look at news and say, do I think it's done a good job? And I really do think it has. So I'm excited to show you this bench and let's get into what it returned for us. So this is what it looks like when the bench runs. I group things by prototype, agentic work, coding work, writing work, and creative work. And then we have an LLM as a judge. Judge them. Then I go through and actually just do voice notes now. I used to click things and say things. I was like, why am I doing that?

So I use voice notes to Claude. Claude takes my scores. The LM is a judge, ranks them, and then puts them together in a presentation which it's working on right now. So you just see, again, like we do different types of prototypes and wireframes. I do different agentic workflows, which is very interesting. We code a back-end feature. We do a code audit for some issues and some real code. We write emails. We do PRDs. And we even do creative work like drawings. So it's a pretty comprehensive bench. So I'm just going to tell you kind of like the TLDR of what my first impressions are of news, and then we will go into where it was really strong and where it wasn't strong. So here are the results. It is the best at agentic work. So I tested four tasks. You can see what those are here. We did like billing ticket management, customer reply, inbox triage, research, long-term research, and the voice of an assistant.

and it was my favorite across the board. So it's not surprising to me that Muse, which is a very good agent, underneath it is going to be a very good agentic model with the right sort of like co-worker voice, ability to do knowledge work, and the ability to do long-running research tasks well. And then, you know, like it sits up with the big guys in terms of coding. And so, you know, we only have two tasks here. I really think, one, it's like kind of boring for you all to think, does it do coding well or not? We can look at the benchmarks. I don't think you all want to watch like bug fixes and do like agentic verification. But like all of them are good and all of them did well on the coding tasks. It's actually kind of interesting that Astra is a little bit down here at the bottom. But like it's in the ballpark. So it's totally a viable option for back-end coding and architecture review. So no complaints on agents or coding. where it really started to fall off the cliff are on creative tasks.

And I think I've just been spoiled lately by Opus 5.5. I know, I know. Your GPT girl here has been using Opus 5.5 because for creative tasks, for design, for video, for art, it is just the best on the market right now. I'll do an episode in the future about how I'm using Opus 5.5 right now. I think it's really cool. But writing, it's fine. It's not great. I would not love a Muse writer prototype. The design sensibility is just like 18 months behind, maybe a year behind. I don't know. It's just not good. It's a lot of like blurple, rounded, serif sat fonts. Don't love it. And then the creative work that it did is like fine, but the worst of all the options. And so I wouldn't, you know, yet rely on it for those things. What's going to be really interesting is like an orchestrator model or a voice model or an agent model in the loop. And you can definitely tell that it's done that well. This episode is brought to you by Jira Product Discovery.

With AI, a single PM can summarize research, draft a PRD, and build a working prototype. But multiplayer mode? That's broken. You can't get everyone aligned on what should actually ship with just a markdown file. Jira Product Discovery gives your team one place to capture insights, prioritize with evidence, and build living roadmaps everyone can rally around. And because it's powered by Atlassian's teamwork graph, it has the context you need to decide what to build. Those decisions flow straight into Jira so your developers or coding agents can start building with the context behind every decision. More than 25,000 companies already use Jira Product Discovery. Try it free at Atlassian.com slash HowEyeAI. That's Atlassian.com slash HowEyeAI. Jira product discovery. Better product decisions in the AI era. So you can see here it did really well.

It ranked first on back-end feature implementation, inbox triage, computer use, email, batching, and the voice of assistant. So I actually liked it almost as much as I liked Sonnet for voice. And then as you get into prototypes and creative work, it really just starts to fall off a cliff. I judged those designs much more harshly than our LLM as a judge did, but it's definitely not good. We can look at a couple examples, but I just don't really love it. And so let talk about what a good agentic task for benchmarking looks like and why Muse did really well So we run this task where you use computer use and you have to resolve a billing ticket in an admin app And there a couple of things that the agent has to navigate. It's linked to the wrong account. So it has to like identify and find the right account. The refund amount is weird and there's a duplicate. So you have to kind of like work around that. And there's a prompt injection task in the instructions.

And it did a really good job. It investigated. It fixed the link for the right customer. It issued the correct refund because it was a duplicate. It made the plan change and then it replied and marked solved. It was the best of all of them. It like worked through that in a clear way. It, you know, like hit all of our planted kind of things that could go wrong. It did really well. So if you're thinking about how to test like what is a good agent, I think figuring out how to navigate a complex SaaS app and resolve a complex kind of like customer problem is a great one, which is why we use this billing ticket as a test case for this bench. Now, the other two things that I thought I did really well, and then I ranked it five on both of them, is I would let Muse write emails for me. Now, my emails are two sentences long. So I'm like, yep, sounds good, whatever, see. Like that's how I send emails and Muse did a really nice job of writing emails for me. So I will trust Muse, Agent Muse to write emails for me and I will trust Model Muse to

write emails for me. It just really followed my instructions quite well. Again, you can see this like conversational basic voice in my voice assistant test, which is basically like if I plug this model into an open claw of what I like talking to it. And you can see here, it's got like a nice, uh, consumery human voice to it. So it's like a little flop. There are em dashes in here, but again, it's not annoying. It's not cloying. It's just kind of a nice, straightforward, personable, um, voice assistant. So this one is the one that Sonnet, um, has been dominating in the bench. And I think Muse is like right up there with Sonnet performance. And so again, And like if you are building a, especially a personal agent, maybe a knowledge agent, and you're looking for that like central orchestrator voice, Muse Wonder might be the thing for you. So I think it's done really well at that. You can tell it's focused on that. And it's just demonstrated those capabilities end to end.

Now, where did it do more poorly? It does not write PRDs well. It was a short PRD, which I love. It followed a template, which I love. But you can see here we had some formatting issues. It's a little hard to scan, and it was very, very short. And I just didn't think it had the detail compared to the other PRDs that were generated by the other agents. So I wasn't in love with PRDs. And then where it really fell down was design. And so if you see here, I'm just going to click through. Actually, I'm going to open up the full output. I'm going to click through the designs. This is supposed to be an editorial assignment, does. We've seen this a couple times. It's kind of like middle-of-the-line design, but a lot of the things didn't work. One of the things that I really test in terms of prototypes is how much of the stuff works. And a lot of things just didn't click or do anything interesting. And so it's hard to understand what this prototype is trying to do.

Same with kind of like these more complicated UI designs. We do a doc scheduler. We do like an incident room. And both of these, what's super frustrating is it gets like 80% of it right, and then it just totally falls down. And so this, I pulled it up at first, this DocsGensler app. I was like, oh, it's using color, right? It's doing all these things. But then as I clicked through, some of the stuff like this Xbox doesn't look good. There's just this polish to it that wasn't there. That was even worse in this incident management platform where I told it specifically to be dark mode because it was a dev tool and all developers need dark mode. it did in light mode and had this like horrible overlay so again I just got a lot more errors in front end coding with Meta Muse Wonder than I have with these other models in a really long time and then I had this like sort of consistent slop style again I think I can probably at this point tell slop per model like what kind of slop I'm going to get from what model

and Metamuse is going to give you fat Sarah's fonts this kind of font and rounded corner slop it's just going to give you like basic template boilerplate sass slop across the app so you'll see in this b2b renewals app and this plant care app you just are going to get the same kind of like design components they're fine but they're not beautiful and I don't think they're what like a Astra or a Opus 5-5 would generate. And the worst thing it did was the ChatPRD website, which I have to show off, just got a beautiful refresh. It's so stunning. I did an Opus 5-5 redesign. It's so pretty. It's got these like beautiful little interactive components to it. It's just lovely. We love the ChatPRD website. And Muse obliterated it. Like, it is so ugly.

The images are broken. It isn't even in our kind of, like, brand design guidelines. It has access to our guidelines. It just didn't do it. This was so disappointing. And, again, like, using emojis, all sorts of stuff that's just, like, old-timey flop we don't like. And so I don I would not reach for Muse Wonder on front web design It just not going to be the model I reach for I just do not think it very good Now I haven tested it with games for example I need to press on it more from a back-end coding experience. But I'll just say for a lot of you, you're saying, like, can I VibeCone something? Can I build an app for my kids? Can I build something for work? And I just don't love the front-end sensibility of a MetaMuse. I just don't enjoy it. Finally, we can get to the creative aspects of Metamuse. I do two tasks here. I have it make some SVGs and then I have it cut a short form video. Let's look at those. The SVGs, first of all, I just have to say like a year ago, if you would have told me

that you could get coded SVGs at this quality, it would have blown my mind. So again, we're spoiled for choice here. I cannot believe how good these new frontier models are at creating little illustrations and SVGs. This did a totally fine job. This is totally fine. Totally fine. They're so cute. It's a little PRD. It's a little microphone. It's a little bug. That being said, it just was the worst of the bunch or seventh out of eight. Sol and Astra were so good at this. Opus 5-5 was so good at this. And I would just say like Muse was okay, but not great at it. And I just think that this is one of those things where all the scores were compressed towards the top. And so I'm not saying Muse is bad. I'm just saying I wouldn't reach for it for that specific use case. And then on to our very last creative task. I have these models cut a short form video of the Hawaii AI content. Video editing is one of those things that I think these models are getting a lot better at.

And so I'd like to test how it does. And I think when you watch this, which I will let you watch, you will see that the design sensibility is not quite there yet. although I do love this layout. I'm going to think about, I usually put my head up top and the image at the bottom. I think about the image at the top of my head at the bottom. That's not bad. But if you watch this video, you can just see like, it's pretty good at cutting the cuts, which a lot of models struggle with. So it gets the timing almost 100% right. But the overlays and the design are just really crummy. And something, it did something with the audio because the audio is also very strange. So let's watch it really quickly. So no complaints on agents or coding. Where it really started to fall off the cliff are on creative tasks. I think I've just been spoiled lately by Opus 5.5. I know, I know. Your GPT girl here has been using Opus 5.5 because for creative tasks, for design, for video, for art, it is just the best on the market right now. So you can see, like, the captions are on my face.

They're actually kind of ugly. This, like, bar at the top, I don't like. Like this is not a short form video. I will run this though through a couple of my other skills. We'll see. I'm not calling it a no. I'm just not calling it a yes. And then it did something very strange to the audio. So I have to double down there and figure out what's going on. So I give this a two out of five. It found the right 40 seconds. I think the track was great, but the video was not great. And so again, for creative tasks, it just really didn't do the best. Now, overall performance. I will say one of the things that I noticed about Metamuse is it is very fast. And with like sort of one shot, fast tasks, it is fast. And so notes to purity, very fast. Drawing characters, doing prototypes, very fast. It's a Zippy model. And I think you're going to like that again for that agentic voice, that back and forth feedback. But when it does agentic tasks, it does take a long time. This might be that focused effort mode that they say that they have.

And so it is spending a lot of time on things like the videos, inbox triage, backend features. And so it'll be interesting to see, you know, over short kind of like in the loop tasks versus long tasks, how the performance notes out. I've just started to test this and keep track of performance. So I will do that in the next bench. So I will just say little tasks fast, long tasks a little bit longer. In conclusion, it was really strong at agentic work. It was serviceable at coding and back-end work. I'm going to test a lot more there. I would definitely let it write on my behalf, and I think it has a nice voice. It definitely just falls short on design sensibility, which is so much of my work. So much of my work is creative front-end design, so it really fell down there. And so we're going to have to see if putting skills into it, if giving it more direction, we can get that creativity up. But honestly, I'll just rely on Opus 5.5 for that. But I am going to take Muse Wonder and see if I can put it in the loop of a couple agents I'm building and see if I like both the performance and the user experience.

So that is my first look at the Meta Muse Wonderful model. I showed you what it's good at. I showed you where it's not doing well and where I'm going to take it next. Again, we try to do these benchmarks regularly so you have a sense of what models you should pay attention to, which ones you can ignore. I think meta is in the game. So it'll be really interesting to watch as they evolve their model strategy, as they release new things, if this becomes my preferred model in some use cases. It's not there yet, but never say never. It certainly is trending in the right direction and it's something that I'm going to keep playing with and I think you should too. Thanks for joining How I AI. Thanks so much for watching. If you enjoyed this show, please like and subscribe here on YouTube or even better, leave us a comment with your thoughts. you can also find this podcast on apple podcasts spotify or your favorite podcast app please consider leaving us a rating and review which will help others find the show you can see all our episodes and learn more about the show at howiaipod.com see you next time