Meta Introduces Muse Wonder Frontier Models With Active Memory
Meta’s Superintelligence Lab has released Muse Wonder and Muse Wonder Full, describing the pair as its first frontier models. The company positions them as larger, more capable successors to Muse Spark, with stronger results on benchmarks and complex tasks including coding.
Muse Wonder is available through the Meta Model API at $4.00 per million input tokens and $20.00 per million output tokens. Muse Wonder Full, meanwhile, will be available only to Muse Code subscribers and will not be served through the API. Both models are also intended to power the Muse agent.
A move beyond the lightweight Muse Spark
Meta released Muse Spark as a lightweight model intended to operate at low latency. The new Wonder models represent a different tier of the company’s AI portfolio: Meta says they are larger and more capable, and its model-card testing places Wonder and Wonder Full at benchmark levels comparable with Fable 5.1 and GPT-6 Astra.
The release follows the path set by Meta’s Muse Spark agentic coding launch as Meta extends its Muse portfolio from agentic coding toward larger models for more demanding work. Meta calls the new pair the next stage of its scaling ladder and links the launch to its pursuit of personal superintelligence.
Memory is central to Meta’s product case
A core focus of the announcement is what Meta calls active memory: the ability to retain and synthesize information supplied by a user over extended periods. The company is targeting assistants that can carry out tasks over days, weeks and months, rather than only respond within a short exchange.
Meta highlighted MemoryBench, a benchmark for continual learning and memory-construction frameworks. In its own testing, Meta says Muse Wonder scored higher on MemoryBench than Fable, Opus and Astra. Research scientist Kavya Kopparapu said the team devoted substantial effort to helping a model remember information important to a user over time.
Training and access boundaries
The models were trained at Meta’s Prometheus data center in central Ohio, which came online earlier this year. They are the first frontier-model release since Scale AI founder Alexandr Wang restructured Meta’s AI efforts in August of last year.
For businesses, the launch creates a practical procurement distinction: Muse Wonder can be evaluated through API pricing and access, while Muse Wonder Full requires a Muse Code subscription. Teams considering long-running coding or assistant workflows should therefore separate the models’ access conditions from Meta’s reported benchmark and memory results when deciding which offering fits their deployment plans.