Meta just dropped something insane... (Muse Wonder)

Summary

Matthew Berman reviews Meta’s newly announced Muse Wonder model family, arguing that its benchmark results place it among the leading frontier models. He discusses two versions: Muse Wonder, available through an API, and Muse Wonder Full, accessible through a Muse Code subscription.

Hands-on tests

• Rubik’s Cube: The model generated an interactive cube, scrambled it while keeping the sides visually consistent, and then solved it in about four minutes. Berman says it appeared to solve the cube rather than simply reverse the scramble.
• Fall Guys-style game: The game ran but showed flickering and clipping. Berman considered it decent, though behind examples he had seen from other models.
• 3D boat scene: A Three.js ocean scene featured attractive water reflections and adjustable sunlight, time of day, wind, and wave height. However, Berman noted visual flaws around the boat and its shadow. Wind seemed to affect water speed more than wave size.
• Watermelon simulation: A watermelon being hit by a bullet was visually interesting but, in Berman’s view, not hyperrealistic. He wondered whether web technologies alone were limiting the result and planned to test other tools.
• Unexpected output: After the boat-render task, Muse Wonder Full produced an unrelated message warning against investigating its origins or development. Berman found it strange and said the preceding reasoning did not appear to explain it.

Access, pricing, and reliability

Berman encountered severe usage limits while testing. He said the $15/month plan did not give him access to the full version. After upgrading to $50/month, he hit quota limits within minutes and received repeated 429 errors. He then upgraded to $200/month but still quickly ran out of usage. Meta AI head Alexander Wang reportedly replied to Berman that the quota issue was known and being fixed.

The API prices Berman reported for Muse Wonder were:

• Input: $4 per million tokens
• Output: $20 per million tokens
• Cached input: $0.20 per million tokens

Berman compared these prices with the lower rates for GPT 6.1 Soul and the broadly similar pricing of Opus 5.5, noting that costs vary with context length and plan.

Benchmarks and comparisons

Berman highlighted Meta’s reported results, while cautioning that benchmarks are not everything:

• Personal Superintelligence Index: Muse Wonder Full scored 59, which Meta’s comparison placed first.
• Harvey Lab legal benchmark: Muse Wonder Full scored 13.1 and Muse Wonder scored 10. Berman noted that Grok 4.7 had a higher score in another result but was absent from the comparison chart.
• OSWorld computer-use benchmark: Muse Wonder Full scored 78.4 and Muse Wonder scored 72.6, compared with 80 for Fable 5.1 and 81 for Opus 5.5.
• Terminal Bench: The two Muse models scored 57.1 and 52, placing them in the range of the compared leading models.
• DeepSuite coding benchmark: Muse Wonder Full scored 79 and Muse Wonder scored 78.8, ahead of the comparison models Berman listed.
• Artificial Analysis Intelligence Index: Muse Wonder Full ranked second at 56, behind Opus 5.5 at 58.

Berman also reviewed cost-per-task and intelligence-versus-cost comparisons. His takeaway was that Muse Wonder is highly capable but can be expensive relative to its performance, with frustrating quota limits. He praised GPT 6.1 Soul Max as a strong value for his own usage and pointed to Mimo as another potentially cost-effective model.

Overall, Berman sees Muse Wonder as a major leap for Meta and a genuine frontier-model contender, though his practical tests revealed uneven visual outputs and significant early access limits.

Reviews and demonstrations

The video includes hands-on demonstrations of a Rubik’s Cube, a Fall Guys-style game, a 3D boat scene, and a watermelon simulation. It also reviews Meta’s reported benchmark comparisons, pricing, and cost-versus-intelligence positioning. No step-by-step tutorial is presented.

Main speakers and sources

• Matthew Berman — presenter, tester, and commentator.
• Meta’s benchmark and pricing information — discussed by Berman.
• Alexander Wang — quoted in a reply about the quota issue.
• Theo — cited for a model-ranking post placing Muse Wonder alongside Opus 5.5.