Tech • AI • Robotics • Game
Muse Wonder: Meta’s new model DESTROYS Claude and GPT-6?
7/10TL;DR
Meta’s M Wonder and M Wonderful appear competitively priced on benchmark tables, but side-by-side coding and simulation tests showed results that lagged well behind Claude Opus 5.5 and often behind Gemini 4 Argon.
KEY POINTS
Benchmark scores raised expectations
M Wonder was highlighted as scoring around 56 on Artificial Analysis, a level presented as comparable to strong frontier coding models while costing less per task than several rivals. Pricing cited in the comparison put it below tools such as Claude Sonnet 5.5 and Opus, creating the impression of unusually high value for money. That combination of benchmark ranking and lower cost set expectations that Meta had produced a major challenger.
Confusion between M Wonder and M Wonderful
The testing quickly ran into a naming problem: M Wonder and M Wonderful were both listed, with the latter appearing stronger on some coding leaderboards. In one benchmark view, M Wonderful ranked near the top of M Code, reportedly in third place. That distinction mattered, because the cheaper and more accessible model did not consistently match the stronger benchmark narrative attached to the premium variant.
Crash simulation exposed major weaknesses
In a car-crash coding test, M Wonderful produced unusual behavior instead of a realistic impact simulation. The virtual car repeatedly tried to avoid the wall or glitched into implausible motion, even after changes to mass, speed, car length, and other parameters. The visuals were at times amusing and conceptually creative, but they did not approach the coherence previously achieved by Claude Opus 5.5 on the same task.
Basic M Wonder looked dated
The standard M Wonder 1.0 result was described as visually primitive, more like a 1990s game or a rough Roblox prototype than a polished browser simulation. Compared with newer frontier models, the output lacked refinement, physical consistency, and interface quality. Even where it functioned, it looked like a large step back from the current top tier.
Higher settings did not solve reliability problems
Moving to extra high settings increased cost without reliably improving quality. One M Wonderful run reportedly took about 1 hour 55 minutes and cost around $11, yet still delivered playback problems and underwhelming output. Another comparison suggested the premium Muse variant remained cheaper than Opus 5.5, but the gap in quality made the savings harder to justify.
Lava lamp test showed partial promise, but not leadership
On a lava-lamp simulation, M Wonderful produced a result that was at least recognizable and more coherent than its weakest outputs. That run reportedly cost about $3 and took 32 minutes, which was notable because it spent longer on this simpler concept than on some crash tests. Still, Claude Opus 5.5 looked far more polished and physically convincing, suggesting that Muse could produce acceptable work but not category-leading work.
Video generation and editing results were disappointing
In generative video and editing-style tests, Muse outputs were judged structurally weak, with poor spacing, weak taste, and inconsistent presentation. A direct comparison with outputs from stronger systems made the gap obvious: the top models felt studio-like and coherent, while Muse looked rough and underdesigned. Even when the task was completed, the finish was not considered competitive with premium alternatives.
Routine coding may still be adequate
Despite the poor showing in ambitious visual tasks, the model was considered capable of handling ordinary engineering requests. For standard product work inside tools such as Cursor, many reasoning models now complete precise, scoped coding tasks competently. On that measure, M Wonder was not uniquely bad; it simply failed to stand out once tasks required taste, surprise, or broad reliability.
Cheaper rivals looked more convincing overall
The broader takeaway favored models that combine strong low-reasoning performance with speed and consistency. GPT-6 Astral, Claude Opus 5 low, and Grok 4.7 low were cited as examples of models that remain highly capable even without expensive extended reasoning. By contrast, Muse appeared optimized to score well on benchmarks while struggling to convert that into superior real-world outputs.
CONCLUSION
The tests suggest Meta’s Muse family may be competent for standard coding, but its headline benchmark strength did not translate into top-tier performance on demanding creative and simulation tasks. For now, lower trust in output quality outweighs the apparent pricing advantage.
Ask a question