Meta Muse Wonder Benchmarks: How It Stacks Up to Opus and Gemini
Meta's Muse Wonder hits 79% on Deep Suite and rivals Claude Opus 5.5 on paper. Here's how the benchmarks compare to real-world coding tests.
What is Meta Muse Wonder and why does it matter?
Muse Wonder is Meta’s newest large language model, and it marks the first time Meta has posted benchmark scores that sit near the top of the closed-model leaderboard alongside Claude Opus 5.5, Claude Sonnet 5.5, and Google’s Gemini 4 Argon. On Deep Suite, a widely used coding benchmark, Muse Wonder scored 79%, reportedly the best mark posted by any top-tier model on that test. It also posted state-of-the-art or near-state-of-the-art results on the Super Intelligence Index, Harvey’s legal benchmark, and a financial analyst research benchmark. For a company that has spent the last couple of years trailing OpenAI, Anthropic, and Google on frontier model quality, that’s a real shift, at least on paper.
TL;DR
- Muse Wonder scored 79% on Deep Suite, which multiple sources describe as the top score among frontier models on that coding benchmark.
- On the Artificial Analysis composite index, Muse Wonder landed above Gemini 4 Argon, GPT 6 Astra, and Claude Fable 5.1, and just two points behind Claude Opus 5.5.
- Pricing runs $4 per million input tokens and $20 per million output tokens, putting it roughly in the same tier as Claude Opus 5.5.
- In hands-on testing, Muse Wonder’s output didn’t match its benchmark ranking. It placed 14th on one creator’s image-generation leaderboard and produced a functional but visually rough browser game compared to rivals.
- Meta also released Muse Spark Open, an open-weight model described as the most capable openweight model out of a US lab, with 800 billion total parameters and 42 billion active at a time (mixture-of-experts architecture).
- Running Muse Spark Open at home isn’t realistic for most people. Even a distilled FP4 version reportedly needs around 700GB of video memory, and higher-precision versions need close to 2TB.
- Several people covering the model, including one reviewer who called it “S tier,” have also flagged that benchmark scores and practical coding output don’t always line up, a recurring problem across this generation of models.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
How does Muse Wonder perform on coding benchmarks?
Deep Suite is the headline number here. A 79% score places Muse Wonder at the top of that particular leaderboard among current frontier models, ahead of what Gemini 4 Argon and GPT 6 Astra posted. Beyond Deep Suite, Muse Wonder also scored well on the Super Intelligence Index (a broad composite benchmark), Harvey’s legal-domain benchmark, and a financial analyst research benchmark, landing at or near state-of-the-art on each.
On Artificial Analysis, which blends multiple benchmarks into a single intelligence score, Muse Wonder ranked above Gemini 4 Argon, above GPT 6 Astra, and above Claude Fable 5.1, a model that drew significant attention a couple of months ago. It sits tied with Claude Sonnet 5.5 and roughly two points behind Claude Opus 5.5, the current top performer on that index. On cost per task, Muse Wonder ranks as the fifth most expensive model tested, cheaper than Anthropic’s other models but still a premium-tier price point.
How does it compare to Opus 5.5 and Gemini 4 Argon in practice?
This is where the picture gets murkier. One reviewer who tracks model releases closely ranked Muse Wonder in “S tier” alongside Opus 5.5, ahead of GPT 6.1 and other current models, based on its benchmark profile. But hands-on testing told a different story. When given the same creative and coding prompts used across other models, including a “build a 3D platformer” style game test, Muse Wonder produced a functional result with correct game mechanics (camera control, jump, sprint) but notably flat graphics, described as a snowman swinging a stick in a sparse environment. By comparison, Gemini 4 Argon’s version of the same test looked more polished, and Claude Opus 5.5 reportedly produced the most visually complete result, followed by GPT 6 Astra.
On a separate image-generation leaderboard where outputs are ranked by an LLM judge, Muse Wonder came in at 14th place, despite its strong benchmark position. The pattern repeats what’s been seen with other heavily benchmarked models this year: strong composite scores don’t always translate into the most impressive or most polished real-world output, particularly for creative and visual tasks.
Is Muse Wonder worth using over Claude Opus 5.5?
On pricing, Muse Wonder and Opus 5.5 land in roughly the same range, Muse Wonder at $4 per million input tokens and $20 per million output tokens. Given that Opus 5.5 outperformed Muse Wonder in direct coding and game-generation tests, at a similar price, the practical case for choosing Muse Wonder over Opus 5.5 right now is thin unless you have a specific reason tied to Meta’s ecosystem or API terms.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Where Muse Wonder’s release matters more is competitive positioning. Meta, Google, and xAI have effectively been competing for a clear third place behind OpenAI and Anthropic, and Muse Wonder’s benchmark scores suggest Meta has closed a meaningful gap. A model with this benchmark profile, had it shipped from Meta even five months ago, would have been treated as a surprising leap forward. Arriving now, alongside Gemini 4 Argon and GPT 6.1, it instead reads as evidence that the frontier is getting crowded, with multiple labs clustered close together rather than one model running away with the top spot.
What is Muse Spark Open, and can you run it yourself?
Alongside Muse Wonder, Meta also released Muse Spark Open, an open-weight model that was previously Meta’s best internal model before Wonder took that title. Meta describes it as the most capable openweight model out of a US lab, a qualified claim given that several Chinese labs have released openweight models that may be more capable overall, but it’s a meaningful marker for openweight development out of the US specifically.
Muse Spark Open uses a mixture-of-experts design with roughly 800 billion total parameters and 42 billion active parameters at any given time. That scale puts real hardware requirements on anyone hoping to run it locally. Even a distilled, lower-precision (FP4) version reportedly needs around 700GB of video memory, and higher-precision variants need close to 2TB. That rules out consumer hardware entirely and points this release toward research labs, cloud providers, and firms with serious infrastructure, or toward the broader ecosystem of smaller distillations the community tends to build once weights are public. Meta reportedly worked with Apple to get the model optimized to run on Apple’s Mac Studio line with M5 Ultra configurations, which narrows the “can I actually run this” question somewhat, though it’s still far from a laptop-friendly release.
Do the benchmarks actually predict real-world performance?
Not reliably, and that’s becoming one of the more consistent observations across recent model releases. Deep Suite, the Super Intelligence Index, and Artificial Analysis composite scores are useful for broad comparisons, but they’re increasingly disconnected from how models perform on open-ended creative or coding tasks that weren’t part of the training or evaluation process. Muse Wonder’s Deep Suite score and its 14th-place finish on a separate visual leaderboard illustrate the gap directly. The same disconnect shows up with Gemini 4 Argon, which some testers rate as visibly stronger at coding and game generation in practice than its composite benchmark position would suggest, and weaker in other areas like raw task completion compared to GPT and Claude models.
The practical takeaway for anyone picking a model for actual work: treat benchmark leaderboards as a first filter, not a final verdict, and test models directly on the kind of task you plan to use them for before committing to one based on a score.
Frequently Asked Questions
What benchmark score did Meta Muse Wonder get on Deep Suite?
Muse Wonder scored 79% on Deep Suite, which several sources describe as the top score among current frontier models on that coding benchmark.
How much does Muse Wonder cost to use via API?
Muse Wonder is priced at $4 per million input tokens and $20 per million output tokens, putting it in roughly the same price range as Claude Opus 5.5.
Is Muse Wonder better than Claude Opus 5.5?
On the Artificial Analysis composite score, Muse Wonder trails Opus 5.5 by about two points, and hands-on coding and game-generation tests showed Opus 5.5 producing more polished results at a similar price.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
Can I run Meta’s Muse models on my own computer?
Not Muse Wonder itself, but Meta also released Muse Spark Open, an openweight model with about 800 billion total parameters (42 billion active). Even a distilled low-precision version needs roughly 700GB of video memory, so it requires server-grade or high-end workstation hardware rather than a typical consumer PC.
Why did Meta’s benchmark scores surprise people?
Meta has trailed OpenAI, Anthropic, and Google on frontier model benchmarks for the past couple of years. Muse Wonder’s scores, landing above Gemini 4 Argon and Claude Fable 5.1 on composite indexes, suggest Meta has closed much of that gap, even though real-world testing shows the results are less consistent than the benchmark ranking implies.
