Meta's Superintelligence Lab announced two new frontier models, Muse Wonder and Muse Wonder Full, on Monday. The release represents the company's first foray into frontier-scale AI, following earlier lightweight offerings. Both models are accessible through Meta's Model API, positioning Meta alongside other industry leaders in large-scale language model deployment.
The earlier Muse Spark model was designed for low-latency, lightweight tasks, whereas the Wonder series targets higher capacity and more complex applications such as software coding. Meta reports that the Wonder models achieve superior benchmark scores and deliver stronger performance on intricate problem-solving scenarios compared with the Spark variant.
According to the model card released by Meta, both Wonder and Wonder Full attain benchmark results that are on par with leading systems such as Fable 5.1 and GPT-6 Astra. The internal evaluation places the new models within the top tier of contemporary large language models, suggesting competitive parity with the most advanced offerings currently available.
Meta makes the Wonder series available through its Model API with a usage fee of four dollars per million input tokens and twenty dollars per million output tokens. This pricing structure aligns closely with the cost of comparable services such as GPT-6 Sol or Opus 5.5, indicating Meta’s intent to compete directly on price as well as performance.
A central focus of the Wonder models is what Meta calls “active memory,” the ability to retain and synthesize user-provided information over extended periods. The company cites superior scores on its MemoryBench benchmark, which tests continual learning and memory construction, claiming the Wonder models outperform Fable, Opus and Astra on this metric. Research scientist Kavya Kopparapu noted in a launch video that Meta “make sure your model can remember.”
Meta trained the Wonder models at its Prometheus data center in central Ohio, a facility that became operational earlier this year. The launch follows a restructuring of Meta’s AI organization led by Scale AI founder Alexandr Wang in August of the previous year. Since that reorganization, Meta has introduced the lightweight Muse Spark model and a Muse agent, but has largely refrained from the high-profile frontier releases typical of competitors such as OpenAI and Anthropic.