Muse Wonder Tested: Payment Failed, Sandbox Failed

Author: Prompt Engineer 48 · 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: Muse Wonder Tested: Payment Failed, Sandbox Failed · Channel: Prompt Engineer 48 · 07:18 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, AI developer tools

Summary

The video tests Meta’s Muse Wonder launch through its benchmarks, playground, API, subscription, and coding agent. The creator reports that the subscription payment initially failed, the terminal ran Muse Spark 1.3 rather than Wonder, and a simple snake-game task took 11 minutes and 40 seconds, with the test left unverified after repeated sandbox errors. In a comparison where both agents were asked only to write the file, Muse took 35 seconds and Claude Code took 19.9 seconds. The creator concludes that the coding agent is unreliable in this Windows test, while noting that they did not verify Wonder’s benchmarks or test its memory and focus features.

Key points

The creator tested Muse in the playground, set up API billing, subscribed to Muse Code, and compared it with Claude Code.
Meta’s published benchmarks showed mixed results: Muse led on some of its listed tests but scored below Claude on the legal benchmark.
Muse Code’s subscription cost 399 rupees per month; the first payment attempt failed and a retry took minutes to process.
The API billing page said payment would be charged when the balance reached 1,000 rupees, so the creator kept API calls small.
The image-generation test produced an acceptable but ordinary-looking image, according to the creator.
The terminal identified the running model as Muse Spark 1.3, not Muse Wonder or Muse Wonderful.
Muse wrote the 191-line snake game file in about a minute and a half, but Windows sandbox errors prevented the test from running.
The task ended after 11 minutes and 40 seconds with verification marked unverified; the file had been written about 10 minutes earlier.
In a write-only comparison, Claude Code took 19.9 seconds and Muse took 35 seconds; Codex produced no result because of a model configuration error.
The creator recommends that users of Claude Code or Codex stay with them, while emphasizing that the test does not establish how Muse Wonder performs on its claimed benchmarks or features.

Chapters

Start	Chapter	Summary
00:00	Test setup and Meta’s claims	The creator introduces the snake-game test, outlines the products tested, and reviews Meta’s models and published benchmark results.
02:08	Subscription, billing, and image test	The creator describes a failed initial subscription payment, explains the API billing threshold, and tests image generation in the playground.
04:10	The snake-game test fails to verify	Muse writes the game file but repeatedly encounters Windows sandbox errors, eventually ending with the test marked unverified.
05:44	Comparison and limitations	The creator compares write-only task times with Claude Code and clarifies that the run used Spark 1.3, not Muse Wonder.
06:47	Verdict and recommendation	The creator gives a negative verdict on the coding agent tested and advises Claude Code or Codex users to stay with their current tools.

Quotes

The other 10 minutes it spent failing to test it. — 00:00

This couldn’t complete payment. Please try a different payment method. — 02:38

model set to muse spark 1.3, not wonder. — 03:40

verification is unverified — 04:42

as a coding tool this is not it. — 07:18

Concepts and entities

Muse Wonder, Muse Wonderful, Muse Spark 1.3, Meta, Meta Model API, Muse Code, Claude Code, Codex, sandbox, RunPod

Full transcript (original language)

00:00 11 minutes and 40 seconds. That is how long Meta’s new coding agent took to make a snake game in one HTML file and tell me it was done. It wrote the file in about a minute and a half. The other 10 minutes it spent failing to test it. And when it finally said done, it admitted the test never ran. I paid for the plan. I tested the playground. I tested the API. And I’m going to tell you exactly where the money and the time went. All right. So, Meta just released a new model family called Muse Wonder. And this is their own launch video. Here is what I did with it.

00:31 One, I tested the model in the playground. Two, I set up API billing on the Meta Model API. Three, I took the Muse code plan, which is 399 rupees a month, so I could run it as a coding agent on my own laptop. And then I compared it with Claude code on the same task. Spoiler, it is not close. This is the blog post published October 12. There are two models. Muse Wonder is on the Meta Model API. And in Muse code, Muse Wonderful is only for Muse code subscribers. Meta says state-of-the-art across

01:03 agents, memory, and coding. And this time there are charts. So, let me read them properly. First chart, the Artificial Analysis Intelligence Index, Claude Opus 5.5 is on top at 58. Muse Wonderful is 56. Muse Wonder, the one in the API, is 52 behind three models sitting at 53. So, on that index, Meta is not first. Then, Meta’s own benchmarks. SWE marathon, Muse Wonder full 57.5%. Claude Opus 5.5 at 46.3. Focus bench, 54.2 against 45.8. Harvey’s

01:36 legal benchmark, 13.1 against 40.2. Memory bench, 85.6 against 81.5. And focus bench and memory bench are labeled internal evaluations. Meta wrote the test and Meta won the test. So, on paper, Meta wins its own exams. I wanted to see if that survives a snake game. Let’s read what they actually promise. Active memory, it keeps contacts across your conversations. Active focus, it keeps working toward a goal, tries other approaches when it hits a blocker, and checks in before important actions. Then domain expertise in law, finance, and

02:08 health, and frontier coding for long-running projects. The demos are Saturday run plan, custom swag from your photos, a wedding seating chart, a 3D kitchen game, and an insurance claim. Now, remember those charts say wonder and wonderful? Hold that thought because that is not the model my terminal told me it was running. Install is one line in PowerShell, and it downloads about 446 MB. Fine, that part worked. Then the plan, Muse code everyday usage, 399 rupees a month. So, I clicked it, and the first thing I got was, “This

02:38 couldn’t complete payment. Please try a different payment method.” So, I switched to my card, got the bank OTP, entered it, and now you can see still processing. It took minutes. I’m not saying this is Meta’s fault for every bank in India, and to be fair, it did go through on the retry, but the first experience of the product was a payment that failed and then hung, and that is before I wrote a single prompt. Next, the Meta model API. You make a project, you create a key, and then billing. This is pay as you go. And here’s the line I want you to look

03:09 at. You’ll pay when your balance reaches 1,000 rupees. So, you’re on a 1,000 rupee meter, which is why I only ran very small calls from here on. Live billing, so I kept it tiny. Playground, image tab, model Muse image 1.0. My prompt even had a typo, generate on purpose, a woman in an ocean asking for help top view. And you can see the result. It is a top-down shot. The hand is up. The face is panicked. The water ripples are fine. It is okay. It is a stock photo. There is nothing here I cannot get from the

03:40 image models I already use. Also, look at the settings. Reasoning on high, image search on, shell on, web search on. For one picture, that is a lot of machinery for an okay image. Now, the agent itself, I type muse, I log in, and look at this line, model set to muse spark 1.3, not wonder. The launch blog is about wonder, and my terminal is running spark 1.3. It loaded 85 skills from my machine, including my Claude code and Codex ones, which is nice. I ask what skills it has,

04:10 it answers. Then the test, can you create a simple snake game in HTML? Nothing fancy, one file. This is the test I run on every agent. Here is what happened. About a minute and a half after my prompt, it wrote snake.html, 191 lines. Good, that is a finished job. But then it decided to smoke test the game with node, and the test never ran. Temp right was blocked, then sandbox hiccup, retrying, then this long red error, Windows elevated unified exact session launcher unavailable, sandbox

04:42 ACL failure. Then shell sandbox is throwing lock errors, retrying once more, then shell execution is down, so I will verify the file by reading it back instead, and then it kept thinking, 11 minutes, 38 seconds of thinking. Then finally, at 11 minutes 40, it said done. And in the same message, it says verification is unverified, the sandbox went down before it could run the test, so it just read the file back, and then it kept double-checking. The file had been written 10 minutes earlier. So I ran a fair test, same laptop, same

05:13 prompt, and I told both not to run it, just write the file. Claude code, 19.9 seconds, muse, 35 seconds, and the file is smaller, almost double. And to be fair to Meta, a one-word prompt takes about the same on both, 11 seconds versus 10.7, so the plain first response is not the problem. Codex errored on this machine with a model config issue, so I have no Codex number, and I’m not going to invent one. The problem is the agent loop, the moment it has to use tools on Windows, it falls over. Now, look at what muse code actually is.

05:44 I ran muse {dash} {dash} help, resume a session, exact for headless runs, skills, plugins, MCP servers, a sandbox command, export a transcript, Claude code has those, Codex has those. It even says it found my Claude code and Codex rules and skills and imported them. So, this is not a new idea. It is the same shape as the tools that already exist with the model behind it that in my run took almost twice as long on the easy task and then lost 10 minutes inside its own sandbox. If you’re going to copy the interface of

06:15 Claude code and Codex, you have to beat them at the one thing that matters, which is finishing the job. On my laptop, it took 11 minutes and ended on unverified. Now, here’s what I could not verify. My plan session ran Spark 1.3, so I never ran Muse wonder or wonderful inside the agent. I did not rerun Meta’s charts and I did not test active memory or active focus, so I cannot tell you wonder loses those benchmarks. I can tell you what I got for my money did not finish a snake game. So, when I say it is bad, I mean the thing I ran on this laptop on this day,

06:47 the install, the payment, the sandbox, and the snake game. Those are all on screen and you can judge them yourself. My verdict, this is bad. I took the plan thinking it would be good and it wasted my time and it wasted my money. Meta is trying to compete with Claude code and Codex and right now it is not even in the same room. They say wonder is state-of-the-art. My terminal ran Spark and for one snake game it took 11 minutes 40, 10 of them fighting its own sandbox before telling me the test was unverified. So, here’s my honest advice to Meta.

07:18 Stop trying to compete on coding agents, take a separate path, do the personal assistant, the memory, the Muse app, the things you actually showed in your demo because as a coding tool this is not it. If you’re on Claude code or Codex, stay there. If you’re looking for cheap and reliable GPUs, I use RunPod and the link is in the description. That is Muse wonder. If you want me to test the next one the same way, subscribe and I will see you in another one.