Meta’s Superintelligence Lab has officially entered the frontier AI market with the release of two new state-of-the-art models, Muse Wonder and Muse Wonder Full. Unveiled on Monday, the new models represent the company’s first major foray into the bleeding edge of artificial intelligence capabilities, following a period focused largely on lightweight, low-latency deployments.
The introduction of the Wonder models marks a sharp pivot from Meta’s previous strategy under the Muse banner. Where the earlier Muse Spark model was explicitly designed to be a lightweight option capable of operating at low latency for everyday tasks, the Wonder models are significantly larger, more computationally intensive, and vastly more capable. According to Meta, these new systems deliver substantially higher benchmark scores and dramatically improved performance on complex reasoning tasks, most notably advanced computer programming and software development.
In its official model card testing, Meta reports that both the standard Wonder and the larger Wonder Full achieve benchmark scores that are directly comparable to industry heavyweights such as Fable 5.1 and GPT-6 Astra. Describing the significance of the release in a company blog post, Meta noted that the models together mark the next step on their scaling ladder and represent a major leap toward achieving personal superintelligence.
This elevated scale and capability are naturally reflected in the commercial pricing structure. Available for enterprise and developer integration through the Meta Model API, the standard Muse Wonder model will cost $4 per million input tokens and $20 per million output tokens. This positions the model at a price point roughly equivalent to competitive alternatives like GPT-6 Sol or Opus 5.5. Meanwhile, the more powerful Muse Wonder Full model occupies a more exclusive tier, available solely to dedicated Muse Code subscribers rather than being served through the standard API. Both models, however, will be integrated directly to power the broader ecosystem of the Muse agent.
In keeping with Meta’s long-term strategic interest in developing sophisticated personal digital assistants capable of executing autonomous tasks over extended periods, the company placed particular emphasis on the models’ memory retention capabilities. Meta highlighted a specialized property it calls "active memory," which allows the systems to retain, recall, and synthesize vast amounts of user information over long stretches of time.
To demonstrate these memory enhancements, Meta pointed to performance on the specialized MemoryBench benchmark, a rigorous evaluation framework designed to test continual learning and complex memory construction. According to the company’s internal testing data, Muse Wonder currently achieves a higher score on MemoryBench than competing models like Fable, Opus, or Astra.
Research scientist Kavya Kopparapu elaborated on this core focus in a video accompanying the official product launch, explaining that the development team dedicated a tremendous amount of effort to ensuring the models can accurately remember what is important to users over days, weeks, months, and beyond.
This release represents a crucial milestone for Meta’s artificial intelligence division, marking its first major frontier model rollout since Scale AI founder Alexandr Wang stepped in to restructure the company’s broader AI efforts in August of last year. In the months following that organizational shift, Meta focused heavily on rolling out the lightweight Muse Spark model and introducing the broader Muse agent ecosystem, while notably avoiding the high-end, state-of-the-art model releases that have defined the public strategies of competitors like OpenAI and Anthropic.
The heavy computational lifting required to train these advanced models was performed at Meta’s state-of-the-art Prometheus data center located in central Ohio, which officially came online earlier this year and boasts some of the highest computing capacities in the industry.