Out of Meta’s Superintelligence Labs on Monday came Muse Wonder and Muse Wonder Full, a first attempt at a frontier-grade model family. Roughly a year has passed since Alexandr Wang, formerly Scale AI’s chief executive, arrived to restructure Meta’s AI work.
Size and reach separate the two. Wonder is priced at $4 per million input tokens and $20 per million output tokens, and it reaches the widest audience through the Meta Model API. The stronger variant, Wonder Full, remains inside Muse Code, the coding subscription, with no public API access.
Both models were trained at a facility in central Ohio: Meta’s Prometheus data center, online since earlier this year. Benchmark results appeared in the model card rather than a technical report, so the training recipe stays undisclosed.
By Meta’s own testing, the models sit near Fable 5.1 and GPT-6 Astra. On the Artificial Analysis Intelligence Index, the larger version edges past Anthropic’s second-best model per Wired, yet Claude Opus 5.5 still leads that table.
Memory received heavy emphasis. Holding a model to a stated goal across long research or coding runs is the job of two mechanisms, Active Memory and Active Focus. Research scientist Kavya Kopparapu said in a launch video that the team spent heavily on recall across days and weeks.
Muse Spark and the Muse agent preceded this release. Refusing hazardous outputs is something Meta says the pair was trained to do, scoring well on safety benchmarks. Independent evaluation of those claims will take time, since evaluation methodology appears in neither the model card nor the blog post.
