Meta releases Muse Wonder and Muse Wonder Full, its first attempt at frontier models

Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full on Monday. They are the company's first attempt at frontier models. Muse Spark was built to be light and fast. The Wonder models are larger, and Meta says they do better on complex tasks such as coding. In Meta's own tests, published in the model card, both score about the same as Fable 5.1 and GPT-6 Astra. Through the Meta Model API, Wonder costs $4 per million input tokens and $20 per million output tokens, which TechCrunch says is roughly what GPT-6 Sol or Opus 5.5 cost. Wonder Full is not available through the API. Developers can only get it with a Muse Code subscription, on the Power Usage plan or a new Maximum Usage plan. Both models will power Meta's Muse agent. They accept text, images, video and audio as input and answer in text. Meta stresses what it calls "active memory," meaning the ability to keep and combine details a user has shared across conversations, and points to its results on the MemoryBench test. Meta calls Wonder its most aligned model so far and says a full safety report is coming. According to Help Net Security, the model card says Meta could not rule out that the models reach its high-risk threshold for chemical and biological misuse.