Meta Debuts Muse Wonder as Its First True Frontier AI Model

Meta has officially entered the frontier AI race with the debut of Muse Wonder and its more powerful sibling, Muse Wonder Full. Developed by the firm’s Superintelligence Lab, these systems represent Meta’s inaugural attempt to challenge the current state-of-the-art benchmarks established by competitors like OpenAI and Anthropic. The new models were trained at the company’s recently opened Prometheus data center in Ohio.

Performance and competitive positioning

Internal testing cited in the model card suggests that the Wonder series performs on par with current high-end architectures such as Fable 5.1 and GPT-6 Astra. While the lighter Muse Spark model prioritized speed, these new iterations are designed for complex reasoning and coding tasks. The Scale Report understands that these models mark a strategic shift in the company’s pursuit of personal superintelligence.

Memory architecture and developer access

Meta is positioning these models as leaders in persistent, long-term task management through a feature dubbed active memory. Research scientist Kavya Kopparapu emphasized that the models were engineered to retain user context over extended periods, an assertion backed by high scores on the MemoryBench evaluation framework.

Pricing and the Muse ecosystem

Access to the standard Muse Wonder model is facilitated via the Meta Model API, priced at $4 per million input tokens and $20 per million output tokens. In contrast, the more capable Wonder Full model is gated exclusively for Muse Code subscribers, reflecting a targeted approach to professional and developer utility. These releases follow a restructuring of the firm's AI division led by Alexandr Wang, the founder of Scale AI, who has been steering the company's research roadmap since August of last year.