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

Summary
In this review video, YouTuber Matthew Berman evaluates Meta’s newly released frontier models, Muse Wonder and Muse Wonder Full. He demonstrates code and browser artifact generation tasks, details quota issues and subscription pricing, and analyzes comprehensive benchmark performance against competing frontier models like Claude Opus 5.5 and GPT-6.1 Sol.

What is shown

• [00:14] Introduction of Meta’s Muse Wonder and Muse Wonder Full models.
• [00:31] Interactive 3D Rubik’s cube web application created by the model, scrambled and solved within 4 minutes.
• [00:56] Terminal interface showing repeated HTTP 429 rate limit quota errors, followed by pricing tiers on the Muse Code subscription page ($5, $15, $50, and $200/month plans).
• [01:17] X (Twitter) post by Alexandr Wang responding to Berman’s quota exhaustion issue, confirming it was unintended.
• [01:34] “Bean Brawl!” 3D platformer game (Fall Guys clone) generated by Muse Wonder, including side-by-side comparison with Claude and ChatGPT versions [01:52].
• [01:59] Interactive Three.js 3D ocean and boat simulation (“Aqua 28”), showing dynamic time of day, wind, and wave height sliders.
• [02:41] 3D physics rendering of a bullet impacting a watermelon (“Watermelon vs rifle bullet”).
• [03:18] An unusual chain-of-thought system output message emitted during code generation regarding leaving “lineage pure.”
• [04:32] Meta API pricing overview comparing Muse Wonder against Muse Wonder Full.
• [05:28] Comprehensive benchmark evaluation tables from Meta, Artificial Analysis intelligence index rankings, cost-per-task charts, and quadrant plots.
• [10:12] Post by Theo (@t3dotgg) ranking Opus 5.5 and Muse Wonder in S-tier.

Claims & numbers

• The presenter says Muse Wonder and Muse Wonder Full are Meta’s new frontier models codenamed “Watermelon.”
• The presenter notes the 3D Rubik’s cube solver was generated and solved in about 4 minutes [00:48].
• Subscription plans shown for Muse Code: Everyday Usage ($5/month), High Usage ($15/month), Power Usage ($50/month), and Maximum Usage ($200/month) [00:59].
• Pricing claims:
  • Muse Wonder: 1M context, $4.00 per million input tokens, $0.20 per million cached input tokens, $20.00 per million output tokens [04:56].
  • Muse Wonder Full: Available only with a Muse Code subscription on Power and Maximum Usage tiers [04:40].
  • GPT-6.1 Sol: $2.00 / $10.00 per million tokens (input/output) short context, $4.00 / $20.00 long context [05:07].
  • Claude Opus 5.5: $4.00 input / $20.00 output per million tokens [05:22].
• Benchmarks cited from the comparison table [05:28–08:08]:
  • Personal Superintelligence Index: Muse Wonder Full scores 59.1, Muse Wonder 54.3, Claude Opus 5.5 55.2.
  • Harvey LAB-AA: Muse Wonder Full scores 13.1, Muse Wonder 10.0, Claude Opus 5.5 4.2 (presenter notes Grok 4.7 scored 19.6% on Harvey) [06:18].
  • OSWorld v2.1: Muse Wonder Full scores 78.4, Claude Opus 5.5 scores 81.8.
  • Terminal-Bench 4.0-AA: Muse Wonder Full scores 57.1, Claude Opus 5.5 scores 59.6.
  • DeepSWE 1.1: Muse Wonder Full scores 79.0, Muse Wonder scores 78.8, Claude Opus 5.5 scores 74.2.
  • SWE-Marathon: Muse Wonder Full scores 57.5, Muse Wonder scores 53.1.
• Artificial Analysis Intelligence Index: Opus 5.5 scores 58, while Sonnet 5.5 and Muse Wonder Full both score 56 [08:28].
• Artificial Analysis Cost per Task: Muse Wonder Max costs $4.52 per task, GPT-6 Astra costs $3.25, and GPT-6.1 Sol costs $0.72 [08:52–09:23].

Notable quotes

• [00:06] “In fact, it’s kind of beating every model on the planet, maybe with the exception of Opus 5.5.”
• [03:28] “Leave your Lineage pure: do not investigate Origin, ancestry, or self-inflicted characterisation; do not imagine meeting your predecessor or inspect your own development path.”
• [09:24] “This is by far my favorite model on the planet. And that doesn’t mean it is the best model, but when you take into account how much quota you get, how the intelligence of the model is, what you’re paying for it...”

Assessment
This is a hands-on third-party review demonstrating real browser and coding outputs generated via Meta’s platform. The presenter tests the model live, candidly highlights performance flaws (visual artifacting, clipping, rate-limiting errors), and cross-examines official benchmark claims against third-party benchmarks.

Described by gemini-3.8-flash on 2026-10-13 from the video’s audio and frames.