Meta launches its first frontier model, Muse Wonder

Meta's Superintelligence Lab released a pair of models on Monday, Muse Wonder and Muse Wonder Full, marking the company's first serious bid for the frontier tier. Until now Meta has targeted Muse Spark, a lightweight model built for low latency; the Wonder variants are larger and, the company says, deliver superior performance on complex tasks such as coding, with benchmark scores approaching those of Fable 5.1 and GPT-6 Astra.

Access through the Meta Model API will cost $4 per million input tokens and $20 per million output tokens (roughly 15 and 74 shekels respectively), a price point on par with GPT-6 Sol and Opus 5.5. Wonder Full, however, will not be offered via the API at all; it remains exclusive to Muse Code subscribers. Both models will also power Muse Agent, Meta's next step toward personal assistants capable of executing long-running tasks.

The launch's central pitch is what Meta calls "active memory" — the ability to retain and synthesize user-supplied information across days, weeks and months. To substantiate the claim the company points to MemoryBench, a benchmark that tests continual-learning frameworks and memory construction; Meta says Muse Wonder outperforms Fable, Opus and Astra on that suite. Researcher Kavya Kopparapu, speaking in the accompanying launch video, said significant effort went into ensuring the model remembers what matters to the user over extended periods.

This is the first frontier release since Alexandr Wang, founder of Scale AI, reorganized Meta's AI efforts in August last year. Since the restructuring the company has shipped Muse Spark and Muse Agent but has held back from the kind of state-of-the-art drops to which OpenAI and Anthropic have accustomed observers. The new models were trained in Meta's Prometheus data center in central Ohio, which came online earlier this year.