Muse Wonder Tops Benchmarks But Fails Lawn Mower Game Test vs Opus 5.5

Meta recently released Muse Wonder and Muse Wonder Full, positioning them as their best models for long-running, multi-step agent tasks. Official benchmark tables show Muse Wonder Full outperforming Claude Opus 5.5 across several professional benchmarks: legal agent, financial analysis, health query, and terminal operation.

However, a real-world test by user Max Blade reveals a stark contrast. Using the same prompt to generate an "AAA-grade lawn mower simulator" with beer and cigar details, Opus 5.5 produced a complete cockpit, lawn, and distant buildings — approaching a playable game — while Muse Wonder Full yielded a low-poly test scene with crude textures. One observer noted: "Feels like one side is AAA/indie game, the other is a Roblox version. Muse is still far from the frontier on 3D tasks."

Cost differences are equally dramatic. The API equivalent cost for Muse was $4.11 versus $99.43 for Opus — roughly 24x cheaper. Yet the test consumed $4.11 of the tester's $50 package, forcing an upgrade to a $200 package to complete the run. This suggests that while Muse appears inexpensive, complex tasks may burn through quota faster than expected.

Muse Code, the coding-oriented product, also drew criticism. A user who tried it immediately unsubscribed, citing CLI-only interface, no VS Code plugin, and inability to handle large text pastes. The verdict: "Worse than Codex; Claude Opus 5.5 remains king."

Despite benchmark dominance, Muse Wonder falls short in practical 3D generation and coding tasks. The article questions in which scenarios Muse Wonder actually achieves state-of-the-art performance. It also references a recent podcast where Mark Zuckerberg and Alexandr Wang discussed their model product philosophy and organizational upgrades.