Meta’s Superintelligence Lab on Monday announced a pair of new state-of-the-art models, the Muse Wonder and Muse Wonder Full, the company’s first foray into frontier models.
While Muse Spark aimed to be a lightweight model that could run with low latency, the Wonder model is larger and more capable, with higher benchmark scores and said to deliver better results on complex tasks like coding.
According to Meta’s tests (listed on the model card), Wonder and Wonder Full achieve benchmark scores comparable to Fable 5.1 and GPT-6 Astra.
Mehta’s blog post about these models says, “Together, they represent the next step on our expansion ladder and a giant leap toward personal superintelligence.”
Its scale is reflected in its pricing. Accessed through the Meta Model API, Wonder costs $4 per million input tokens and $20 per million output tokens, making it roughly comparable in cost to GPT-6 Sol or Opus 5.5.
In particular, Wonder Full is only available to Muse Code subscribers and is not offered through the API. Both models are also used to run the Muse agent.
In line with Meta’s interest in developing personal assistants that can perform tasks over long periods of time, the company placed particular emphasis on the new model’s ability to retain and synthesize information provided by the user, a property Meta refers to as “active memory.”
To achieve better recall, Meta highlighted the MemoryBench benchmark, which challenges continuous learning and memory building frameworks. According to the company’s tests, Muse Wonder scored higher on MemoryBench than Fable, Opus, and Astra.
“We’ve put a lot of effort into making sure our models remember what’s important to you over days, weeks, months, and even beyond,” research scientist Kavya Kopparapu said in a video accompanying the announcement.
These new models are the first Frontier model releases since Scale AI founder Alexandr Wang restructured Meta’s AI efforts last August. Since then, Meta has released lightweight Muse Spark models and Muse agents, but has avoided releasing the cutting-edge models that OpenAI and Anthropic are known for.
The model was trained at Meta’s Prometheus data center in central Ohio, which went live earlier this year, according to the post.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
