Meta Muse Wonder Is HERE – Is THIS an Opus Competitor?

Summary
In this video, creator Bijan Bowen tests and reviews Meta’s newly released frontier models, Muse Wonder and Muse Wonder Full (codenamed Watermelon), released on October 12, 2026. He evaluates their capabilities against top competing models like Claude Opus 5.5 and GPT-6.1 across complex coding, browser OS creation, Godot game generation, robotics tasks, and 3D web applications. While initially frustrated by a severe subscription quota bug on Meta’s Muse Code tiers, he finds the models—especially Muse Wonder Full with reference prompts—capable of impressive long-horizon software and game generation.

What is shown

• Meta Blog Post & Architecture Benchmarks [00:10]: Bowen reviews the official release announcement and benchmark tables comparing Muse Wonder / Muse Wonder Full against Claude Opus 5.5, Claude Fable 5.1, and GPT-6 Astra across the Artificial Analysis Intelligence Index, DeepSWE, and SWE-Marathon.
• Quota Glitch & Plan Upgrade [01:40, 10:37]: Bowen demonstrates how two runs quickly consumed his quota on the $50/month Power Usage plan, showing Alexandr Wang’s X post addressing the bug before upgrading to the $200/month Maximum Usage plan.
• NovaOS Browser Desktop Environment [04:54]: Testing a generated single-file browser operating system featuring dynamic wallpapers, a command center voice assistant (“Nova”), an interactive paint app, a generative synthesizer, and embedded games.
• Skatey NYC 2001 (C++ & raylib) [09:07]: Testing a generated standalone 3D skateboarding game featuring street environments, pedestrian interactions, and board mechanics.
• Subway FPS WebGL Game [11:13]: Playtesting a wave-based 3D first-person shooter in Three.js where players fight zombies in a subway station and board moving trains.
• Peri Desktop Robot Interlude [13:40]: Bowen demonstrates “Peri”, his open-source, 3D-printed desktop companion robot running local/cloud voice models.
• Godot / Blender Backyard Pool Party & Brawl [15:39, 18:20]: Evaluating muse-wonder-1.0 generating a Godot diving game, then iteratively pivoting it into a 4-player arena brawler (“Backyard Brawl”).
• Physical Robot Arm Manipulation Test [19:57]: Running a vision-guided robotic arm task attempting to grasp and relocate a toy pickup truck using static single-camera visual feedback.
• Block Party Skate NYC 2002 (Iterative Redesign) [21:52]: Giving Muse Wonder Full a reference photo to remake the skate game, producing an upgraded city park with moving cars, other skaters, and a live picture-in-picture broadcast camera feed.
• Slappis Watch Co. Front-End 3D Site [24:49]: Inspecting a single-file luxury watch showcase with an interactive exploded 3D component view.
• Subway FPS Visual Upgrade [26:16]: Testing an upgraded subway game iteration with enhanced lighting, train arrival boards, blood splatters, and dynamic subway train strap physics.

Claims & numbers

• The presenter notes Muse Wonder was codenamed “Watermelon” during teasing on X [00:19].
• The presenter notes Meta states the models were trained at its gigawatt-scale Prometheus data center [00:30].
• The presenter notes pricing for Muse Wonder API is $4.00 per million input tokens and $20.00 per million output tokens ($0.20 per million cached input) [04:02], matching Anthropic’s Claude Opus 5.5 pricing [04:12].
• The presenter states Muse Wonder Full is accessible via the Muse Code plan subscriptions ($50/month Power Usage or $200/month Maximum Usage) [01:27, 10:37].
• In the Godot pool party test, muse-wonder-1.0 took 9 minutes and 28 seconds on its first attempt [15:19] and 16 minutes and 1 second on the visual overhaul [17:10].
• The presenter notes the robot arm test ran for about 1 hour and 15 minutes before being stopped due to depth-perception limitations [20:42].

Notable quotes

• [00:08] “Meta has come out with their Fable competitor, which I think now would be an Opus competitor, cuz I think Opus beats Fable on a lot of things.”
• [04:01] “It is $4 per million in and $20 per million out. This is expensive, and what is the comparable pricing for this?”
• [22:00] “This right here is a significant, significant, significant improvement and I am rather pleased with what I’m seeing.”

Assessment
This is a genuine, hands-on independent review and live evaluation of Meta’s Muse Wonder model family. Bowen runs authentic terminal sessions and playtests unedited game builds, acknowledging real-time service bugs (such as the early token quota miscalculation) as well as areas where the models struggle, such as single-camera robotic spatial reasoning.

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