Meta’s Superintelligence Lab has introduced its first frontier models, Muse Wonder and Muse Wonder Full, marking a significant advancement beyond its earlier Muse Spark model. While Muse Spark was designed for lightweight, low-latency performance, the Wonder models are larger and deliver enhanced capabilities, including superior benchmark scores and improved handling of complex tasks such as coding. These new models reflect Meta’s ambitions to develop powerful AI that approaches personal superintelligence.

The Muse Wonder models achieve benchmark performance comparable to leading competitors like Fable 5.1 and GPT-6 Astra, and will be accessible via the Meta Model API. Pricing for Muse Wonder is set at $4 per million tokens for input and $20 per million tokens for output, aligning with costs for similar frontier models like GPT-6 Sol and Opus 5.5. However, Muse Wonder Full will be exclusive to Muse Code subscribers and not available through the API. Both models will also underpin Meta’s Muse agent, a smart assistant designed to perform extended, complex user tasks.

A key focus in developing these models has been building “active memory,” Meta’s term for an AI’s ability to retain and synthesize user-specific information over time. Meta highlighted that Muse Wonder outperforms other models on the MemoryBench benchmark, which tests continual learning and memory construction. This enhancement aims to enable AI systems that remember important details over days, weeks, or even longer, enhancing their usefulness as personal assistants.

These frontier models represent Meta’s first such release since the restructuring of its AI division by Scale AI founder Alexandr Wang in August 2025. Training took place in Meta’s Prometheus data center in Ohio, which began operations earlier this year. The launch of Muse Wonder and Muse Wonder Full signals Meta’s renewed push into the cutting edge of AI, balancing increased power and memory capabilities while targeting practical applications in personal assistant technology.