Muse Wonder from Meta | First Impressions

Summary

Peter Gostev, AI Capability Lead at Arena, gives his first impressions and comparative analysis of Meta’s newly released frontier model, Muse Wonder (specifically testing Muse Wonder Max). He evaluates its performance across various agentic code generation tasks, comparing its outputs, visual aesthetic, and execution costs side-by-side against models such as GPT-6 Astra, GPT-6.1 Sol, GPT-5.6 Sol, Claude Opus 5.5, Claude Sonnet 5.5, and Kimi K3.

What is shown

• [00:11] Meta’s benchmark comparison table displaying scores across Personal Superintelligence Index, Harvey LAB-AA, Vals Finance Agent V2, HealthBench Pro, AutomationBench-AA, GDP.pdf, JobBench, OSWorld V2.1, MemoryBench, Terminal-Bench 4.0-AA, DeepSWE 1.1, and SWE-Marathon for Muse Wonder Full, Muse Wonder, Claude Fable 5.1, GPT-6 Astra, and Claude Opus 5.5.
• [00:32] Interactive 3D web model of a Giant Pacific Octopus generated by Muse Wonder Max ($4.21), compared to higher-fidelity outputs from GPT-6 Astra Max ($7.69) [01:06], Kimi K3 ($1.92) [01:55], GPT-6 Luna Max ($0.04) [02:12], and GPT-5.6 Sol [02:45].
• [03:10] 3D Humpback Whale breach simulation generated by Muse Wonder Max ($1.11), compared with generations from Kimi K3 ($2.48) [03:17], Mistral Large 4 ($1.05) [03:38], and GPT-5.6 Sol ($1.47) [03:45].
• [04:06] Procedural 3D scene of the White House and President’s Park by Muse Wonder Max ($1.69), contrasted with GPT-6.1 Sol Max ($3.76) [04:28] and Claude Sonnet 5.5 High ($4.02) [04:55].
• [05:56] “The Lost Valley” 3D dinosaur sanctuary generated by Muse Wonder Max ($1.22), compared with GPT-6.1 Sol Max ($2.55) [06:19], Claude Sonnet 5.5 High ($5.66) [06:46], and Kimi K3 ($0.61) [07:03].
• [07:51] 2D canvas animated scroll story (“The Great Wall Unrolls”) created by Muse Wonder Max ($0.38), contrasted with GPT-6 Astra Max ($2.99) [08:40], GPT-6.1 Sol Max ($0.65) [09:11], and Claude Sonnet 5.5 High ($0.96) [09:35].
• [10:37] 2D canvas animation of Pompeii fresco generated by Muse Wonder Max ($0.47) and GPT-6.1 Sol Max ($0.46) [10:50].
• [11:22] Front-end interactive race dashboard (“Velora Grand Prix”) created by Muse Wonder Max ($1.07), compared with Claude Opus 5.5 Max ($16.36) [11:35], GPT-6 Astra Max ($7.02) [13:27], and Kimi K3 ($3.17) [14:24].
• [15:15] Optical art exhibition website (“Interference”) created by Muse Wonder Max ($0.65) and Claude Opus 5.5 Max ($11.24) [15:52].
• [16:42] Interactive aquarium website (“Lumenmere”) built by Muse Wonder Max ($2.01) compared with Claude Fable 5.1 Max ($19.19) [16:43], GPT-6.1 Sol Max ($2.18) [17:22], and Kimi K3 ($4.81) [17:38].
• [18:22] Residential building floor plan and apartment selector website (“The Kiln”) by Muse Wonder Max ($0.82) compared to GPT-6.1 Sol Max ($1.15) [18:56] and Claude Sonnet 5.5 High ($2.50) [19:29].
• [20:52] Singapore AI conference spatial-planning dashboard generated by Muse Wonder Max ($2.11) versus GPT-6 Astra Max ($14.85) [21:05] and GPT-6.1 Sol Max ($2.41) [21:11].
• [21:32] 3D Acropolis model generated by Muse Wonder Max ($0.95), compared with GPT-6.1 Sol Max ($2.35) [21:43] and Claude Sonnet 5.5 High ($1.12) [21:53].
• [22:54] Arena Agent leaderboard overview, capabilities breakdown, and Pareto frontier performance-vs-cost charts.

Claims & numbers

• Meta claimed Muse Wonder achieves Claude Opus 5.5 or GPT-6 Astra level performance [00:06].
• Meta benchmark table reports Personal Superintelligence Index at 59.1 for Muse Wonder Full and 54.3 for Muse Wonder, compared to 49.9 for Claude Fable 5.1, 54.5 for GPT-6 Astra, and 55.2 for Claude Opus 5.5 [00:11].
• Terminal-Bench 4.0-AA reported at 57.1 for Muse Wonder Full and 52.0 for Muse Wonder [00:18].
• DeepSWE 1.1 reported at 79.0 for Muse Wonder Full and 78.8 for Muse Wonder [00:18].
• MemoryBench context score is listed at 85.5 for Muse Wonder Full and 85.6 for Muse Wonder [00:18].
• The presenter notes that generation costs for Muse Wonder Max consistently land around $0.40 to $4.20 per project, which is substantially cheaper than top-tier flagship runs like Claude Opus 5.5 Max ($11–$16+) or GPT-6 Astra Max ($7–$15), but typically more expensive than lower-tier models like GPT-6 Luna Max ($0.04) or basic Kimi/Mistral passes [01:30, 02:27, 12:43].
• The presenter finds that in practical agentic web development and 3D modeling tasks, Muse Wonder Max feels closer to GPT-5.6 Sol quality rather than matching the current frontier of GPT-6 Astra, GPT-6.1 Sol, or Claude Opus 5.5 [03:51, 05:47, 22:40].

Notable quotes

• “Today we have an important model from Meta... today they claim they have a Opus 5.5 level model or Astra level model.” [00:00]
• “I feel like it’s about 5.6 Sol model... that’s my current thinking.” [03:51]
• “Looking at the benchmarks, I genuinely thought there’ll be much, much more impressive set of generations that we’re going to get.” [20:41]

Assessment

This is an authentic third-party review and hands-on evaluation by an industry practitioner running identical prompts across a standardized multi-model testbed. The generations, UI interfaces, execution prices, and interactive code artifacts are demonstrated live in browser environments without deceptive cuts.

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