Meta launches Muse Wonder, its first frontier AI model

Meta's Superintelligence Lab unveiled its initial frontier AI systems, introducing Muse Wonder and Muse Wonder Full. Departing from the company's previously released lightweight Muse Spark system, these larger architectures target complex computational demands, including advanced software coding. Built at Meta's recently activated Prometheus data center in central Ohio, the models represent the firm's first major high-capability release since Scale AI founder Alexandr Wang reorganized Meta's artificial intelligence division last August.

In internal evaluations, Meta reported that the new models deliver performance on par with rival offerings such as Fable 5.1 and GPT-6 Astra. A central focus of the release is a capability termed active memory, designed to allow long-term tracking and synthesis of user data over extended periods. To evaluate this feature, Meta tested the systems against MemoryBench—a benchmark measuring continual learning—where Muse Wonder outpaced competitors including Fable, Opus, and Astra. Research scientist Kavya Kopparapu emphasized that developers directed substantial resources toward ensuring the models reliably retain vital user context over months.

Access to the technology varies by model tier and deployment method. Standard Muse Wonder is available via the Meta Model API at $4 per million input tokens and $20 per million output tokens, aligning its operational costs with existing market alternatives like GPT-6 Sol and Opus 5.5. Meanwhile, the higher-tier Muse Wonder Full will be offered exclusively to Muse Code subscribers rather than through the API. Both releases will also power the broader Muse agent ecosystem.

For developers and AI practitioners, Meta's expansion into frontier-scale systems signals a major shift toward persistent, memory-driven virtual assistants. By offering competitive benchmark performance alongside specialized active memory capabilities, these systems provide builders with new options for complex workflows that require historical user context alongside high-end coding ability.