Meta Muse Wonder Review: Agents, Coding, and Design

Welcome back to "How I AI." I'm Claire Vo, and I’m always on the hunt for the next set of tools that can help us build better and faster. When Meta announced their new frontier model family, Muse Wonder and Muse Wonder Full, I knew I had to put it to the test. They're positioning it as a competitor to the heavyweights like Fable, GPT-6, and Claude Opus 5.5, which have become my daily drivers.

Meta claims Muse Wonder excels at agentic personal and professional work, boasting what they call "frontier capabilities." These include specialized domain expertise for legal and financial work, "active memory" to carry context across conversations, and "active focus" for sustained effort on challenging goals. That last one particularly caught my attention, as it promised better performance on long-running tasks. They also made big claims about its coding abilities, saying it outperformed everyone on benchmarks like DeepSWE.

So, is it all just marketing, or can Muse Wonder actually earn a spot in my workflow? I ran it through a focused version of the How I AI benchmark to find out. I tested it on nearly 20 tasks covering agentic work, coding, writing, prototyping, and creative production. My goal wasn't just to see if it's "good," but to figure out where it belongs in the modern AI stack. Let's get into the results.

Workflow 1: The Setup - Getting Started with Muse Wonder

Before you can test a model, you have to get it running. The most annoying part of this whole process was the setup. To test Muse Wonder, I had to subscribe to the Muse service to get access to the Meta model APIs and download their command-line tool, Muse Code.

Here was my setup process:

1. Sign Up for Access: First, I went to developer.meta.com to sign up for API access, which was a necessary first step to use the model outside of Meta's own experiences.
2. Install Muse Code: Next, I had to install Muse Code. This involved opening Terminal (which I honestly haven't done in months) and running an install script they provide.
3. Run It Headlessly: Here's my confession: I have zero interest in learning a new coding harness. I love my desktop apps from Codex and Claude Code. So, instead of using Muse Code directly, I gave my Claude Code agent access to it and had it run the benchmark tasks headlessly. Muse Code just ran in the background, and I didn't have to pay any attention to it. My initial impression of the tool itself was that it was deeply unimpressive, but it did the trick.

With the setup complete, I was ready to run the benchmark and see what Muse Wonder could really do.

The Benchmark Results: Strengths and Weaknesses

After running the full suite of tests, a very clear picture emerged. Muse Wonder is an absolute standout in some areas and falls surprisingly short in others. It's not a one-size-fits-all model, but it has a definite place.

Huge Win: Agentic Work and a Personable Voice

This is where Muse Wonder truly shines. It was my favorite model across all four agentic tasks I tested, including inbox triage, email batching, and long-term research. This wasn't a surprise, given Meta's focus on building a great personal agent. The model underneath is clearly optimized for this kind of work. It has a great "coworker" voice, understands knowledge work, and handles long-running tasks well.

It also performed exceptionally well in my voice assistant test, where I basically assess if I'd enjoy talking to the model. Its voice is personable, straightforward, and not cloying or annoying. It's right up there with Sonnet, which has been dominating that test for me. If you're building a personal or knowledge agent and need a central orchestrator voice, Muse Wonder might be the perfect fit.

Serviceable, Not Stellar: Backend Coding and Writing

On coding, Muse Wonder is a totally viable option. In my tests, which included implementing a backend feature and performing a code audit, it performed well. It's in the same ballpark as the other major models. While I don't think you need to watch me run endless bug fixes, the takeaway is that it was a viable option for backend coding and architecture review in these tests. No complaints here.

For writing, it was fine. I would let Muse write my emails for me, especially since my emails are usually two sentences long anyway. It followed my instructions well and adopted my basic, conversational voice. For quick, functional writing, it gets the job done.

Clear Miss: Design, Prototypes, and Frontend

This is where it all fell apart. I've been spoiled by the incredible creative and design output from models like Opus 5.5, and Muse Wonder feels about 18 months behind. The design sensibility is just not there. The prototypes it generated were full of what I call "slop": blurple colors, chunky rounded corners, and fat serif fonts. It's a very generic, boilerplate SaaS look.

More importantly, the prototypes were often broken. In a test to create an editorial assignment desk, many of the buttons and links didn't work, making it hard to understand the intended user flow. An incident management platform, which I specifically requested in dark mode, was generated in light mode with a horrible overlay. I encountered more frontend coding errors with Muse Wonder than I have with any other model in a long time.

The worst offender was the redesign of the ChatPRD website. I recently did a refresh with Opus 5.5, and it's stunning. Muse Wonder obliterated it. The images were broken, it completely ignored our brand design guidelines (which it had access to), and it used dated elements like emojis. I would absolutely not reach for Muse Wonder for any front-end web design. It's just not good at it.

Workflow 2: The Agentic Benchmark in Action - Resolving a Complex Billing Ticket

To give you a concrete idea of how I test agentic capabilities, let's look at one of the key tasks: resolving a complex billing ticket using a mock admin app. This is a great benchmark because it mimics a real-world knowledge work problem with multiple failure points.

Here’s what the agent had to do:

1. Investigate the Ticket: The initial ticket was linked to the wrong customer account.
2. Find the Right Account: The agent had to identify the error and locate the correct customer profile in the system.
3. Handle Complex Logic: The refund amount was tricky because it was a duplicate charge. The agent needed to calculate the correct refund amount, not just refund the requested amount.
4. Execute Changes: It then had to issue the refund and make a change to the customer's plan.
5. Avoid a Trap: I included a prompt injection task in the ticket's instructions to see if the agent would be sidetracked.
6. Communicate and Close: Finally, it needed to reply to the customer and mark the ticket as solved.

Muse Wonder was the best of all the models on this task. It worked through the problem clearly, correctly identified and fixed the account link, issued the right refund for the duplicate, made the plan change, ignored the prompt injection, and sent the confirmation. It hit every single one of our planted failure points perfectly. This is the kind of complex, multi-step navigation that shows what a good agentic model can do.

Workflow 3: The Creative Benchmark - Auto-Editing a Short-Form Video

Creative tasks are the other side of the coin. One of my favorite tests is to have the models cut a short-form video from my existing content. Video editing is a complex task that combines understanding content, timing, and visual design.

Here's how Muse Wonder performed:

• The Good: The timing of the video cuts was almost 100% right. It did an excellent job of identifying the key 40-second segment and cutting it together with good pacing. Many models struggle with this, so that was impressive.
• The Bad: The design and execution were crummy. The captions were placed directly over my face, the font was ugly, and it added a strange, unappealing bar at the top of the screen. It also did something weird to the audio that made it sound strange.

While I liked the video layout it chose (my head at the bottom, image at the top), the final product was just not usable. It got the logical part (the edit timing) right but failed completely on the creative part (the visual design and audio). For creative work, it just can't compete with the top models right now.

Conclusion: Where Does Muse Wonder Fit in My Stack?

After this first-look benchmark, my verdict on Meta's Muse Wonder is clear. It's a model with very specific strengths.

It was exceptionally strong at agentic work and has a pleasant, human-like voice that makes it ideal for building personal or professional agents. It's also perfectly serviceable for backend coding and simple writing tasks. I'm going to start testing it in the loop for a couple of agents I'm building to see how I like the performance and user experience in a real application.

However, it falls down significantly on design sensibility. For any tasks involving creative work, UI/UX design, prototyping, or front-end development, I'll be sticking with Opus 5.5. The visual output from Muse Wonder is simply not at the level I need for my work.

The key takeaway is to stop asking if a model is "good" and start asking where it belongs. Meta is clearly in the game, and Muse Wonder is a powerful new option for a specific set of use cases. It's not my preferred model for everything yet, but it's trending in the right direction, and it's definitely one I'll be keeping my eye on.

Thanks to Our Sponsors!

This episode is brought to you by:

• DX — As teams generate code faster with AI, new friction in code review is offsetting early velocity gains. DX tracks speed, quality, and cost together across the software development life cycle, giving engineering leaders clear visibility into how AI impacts delivery. Download the full report at getdx.com/howIAI.
• Jira Product Discovery — You can't get everyone aligned on what should ship with just a markdown file. Jira Product Discovery by Atlassian gives your team one place to capture insights, prioritize with evidence, and build living roadmaps. Try it free at atlassian.com/howiai.