Meta Enters Frontier AI Race with Muse Wonder Models

Meta’s Superintelligence Lab launched the new models on Monday, positioning them as the company’s most capable offerings to date. Internal testing indicates that Wonder and Wonder Full reach benchmark parity with Fable 5.1 and GPT-6 Astra. The models were trained at Meta’s Prometheus data center in central Ohio, marking the first major frontier-class release since the restructuring of the company’s AI division under Alexandr Wang last August.

The architecture focuses heavily on what Meta describes as "active memory," an ability to retain and synthesize user information over extended periods. Research scientist Kavya Kopparapu noted that the team prioritized long-term recall, allowing the models to remember user-specific details across weeks or months. This capability is verified through the MemoryBench test, where Meta claims its new models outperform Fable, Opus, and Astra.

Accessibility varies by tier: the standard Wonder model is available via the Meta Model API at $4.00 per million input tokens and $20.00 per million output tokens, placing it in a similar price bracket to GPT-6 Sol or Opus 5.5. Conversely, the more powerful Muse Wonder Full is restricted exclusively to Muse Code subscribers. Both versions will serve as the engine for the existing Muse agent.