Meta’s Superintelligence Lab recently introduced two advanced models, Muse Wonder and Muse Wonder Full, marking the company’s initial venture into frontier AI development.
While Muse Spark was designed for speed and efficiency as a lightweight model, the Wonder series offers greater capacity and performance, demonstrating superior benchmark results and improved handling of intricate tasks like coding.
Based on Meta’s internal evaluations, detailed in the model card, Wonder and Wonder Full achieve benchmark scores comparable to Fable 5.1 and GPT-6 Astra.
Meta’s official blog post states that these models collectively represent a significant advancement in their scaling efforts and a major step toward achieving personal superintelligence.
This increased scale is reflected in their cost: accessed via the Meta Model API, Wonder will be priced at $4.00 per million input tokens and $20.00 per million output tokens, aligning its cost with models like GPT-6 Sol or Opus 5.5.
It is important to note that Wonder Full will exclusively be available to Muse Code subscribers and will not be accessible through the API. Both models are also slated to power the Muse agent.
In line with Meta’s strategy to develop personal assistants capable of sustained task performance, the company emphasized the new models’ enhanced ability to retain and synthesize user-provided information, a feature they call “active memory.”
To demonstrate improved recall, Meta highlighted the MemoryBench benchmark, designed to test continual learning and memory construction frameworks. According to the company’s tests, Muse Wonder now surpasses Fable, Opus, and Astra in MemoryBench scores.
“We dedicated considerable effort to ensuring your model can recall important details over extended periods – days, weeks, months, and beyond,” stated research scientist Kavya Kopparapu in a video accompanying the launch.
These new models represent the first frontier model release since Scale AI founder Alexandr Wang reorganized Meta’s AI initiatives in August of last year. Since that time, Meta had previously launched its lightweight Muse Spark model and Muse agent, but had refrained from releasing state-of-the-art models comparable to those from OpenAI and Anthropic.
The models were reportedly trained at Meta’s Prometheus data center in central Ohio, which commenced operations earlier this year.