Summary
Meta Superintelligence Labs has released Muse Wonder and Muse Wonder Full, which the company calls its first frontier-class models. This video from TheAIGRID breaks down what Meta announced on October 12th, how the two models differ, and whether the claims hold up.
Muse Wonder is available through the Meta model API and inside Muse Code, Meta’s terminal-based coding agent. The larger Muse Wonder Full is reachable only through Muse Code subscriptions on the $50 or $200 plans. Both were trained at Meta’s gigawatt-scale Prometheus data center in Ohio. Meta pitches the models around personal and professional agents, with two headline features: active memory, which retains context across days and months, and active focus, which trains the model to persist through blockers instead of giving up. Demos include a playable 3D kitchen game, a personalized running plan built from map data, and a wedding seating chart with complicated constraints.
The video also covers the economics behind the launch, including Mark Zuckerberg’s comments on the cost of running gigawatt clusters, the Hyperion cluster in Louisiana, and Meta’s pace of six releases in six months. It closes with the backlash: many viewers suspect the model is benchmaxed, pointing to early community demos, including a design benchmark run by Bridgemind, that did not match Meta’s claims.
📺 Source: TheAIGRID · Published October 13, 2026
🏷️ Format: News Analysis
