Meta Muse Wonder tested: strong benchmarks, disappointing code

Robbert van Empel published a practical review of Meta Muse Wonder on October 13, 2026, testing whether its strong intelligence benchmark results translate into useful generated software. The video examines a Wikipedia page redesign and an interactive LEGO-style South Park world. Both outputs contain working elements, but also visible problems that leave the reviewer disappointed.

The review follows Meta’s October 12 announcement of Muse Wonder and Muse Wonder Full. Meta makes Wonder available through its Model API and Muse Code, while Full is available to Muse Code subscribers. That distinction matters: Van Empel’s own coding tests use Muse Wonder through an API, so their results should not be presented as a direct assessment of Wonder Full.

The first assignment redesigns Wikipedia’s page about toilet paper orientation. Van Empel likes the resulting visual style, but shows an awkwardly positioned dark-mode button and white text disappearing against a white background. The interactive illustration allows some actions, including switching orientation and tearing a sheet, yet its presentation falls short of his expectations. In the benchmark ranking displayed in the video, this output places eleventh.

The second assignment asks for an explorable world with LEGO-style construction and South Park-inspired characters. Walking and jumping work, but much of the environment does not resemble LEGO bricks. The demonstration shows overlapping objects, movement through buildings and a grabbed object remaining suspended after release. A dialogue prompt tells the player to press E, but the conversation does not continue. This result places ninth in the displayed test ranking.

The video also discusses Artificial Analysis intelligence scores, reported task costs and community demonstrations. Those comparisons provide context, but a benchmark score or another creator’s example cannot guarantee the outcome of a particular coding prompt. Van Empel additionally reports encountering an age-confirmation request before API use; the video documents his experience without establishing that every access route has the same requirement.

For makers and businesses, the practical lesson is to inspect generated software beyond its first impression: readability, controls, collisions and dialogue all affect usability. These two examples establish specific shortcomings in the demonstrated outputs, rather than a comprehensive measure of the model. Despite the disappointing results, Van Empel welcomes Meta’s additional competition in advanced AI.