Meta Enters the Frontier AI Race With Muse Wonder Models and ‘Active Memory’

Meta’s Superintelligence Lab has officially entered the frontier artificial intelligence race, releasing a pair of powerful new models on Monday named Muse Wonder and Muse Wonder Full. This major release represents the technology giant’s first official foray into high-end frontier models, signaling a significant evolution in its long-term artificial intelligence strategy and its ambitions for personal superintelligence.

The newly unveiled models follow the earlier rollout of Muse Spark, a lightweight model explicitly designed for low-latency tasks and basic operational efficiency. In contrast, the Wonder models are considerably larger, more capable, and engineered to handle much heavier computational workloads. According to Meta, these new systems deliver significantly higher benchmark scores and showcase vastly improved results on complex reasoning tasks, most notably advanced computer programming and software coding.

In internal testing detailed in the official model card, both the standard Wonder and the larger Wonder Full achieve competitive benchmark scores that place them directly alongside other industry-leading systems, specifically rivaling Fable 5.1 and GPT-6 Astra. Company executives and researchers are framing this launch as a critical milestone in Meta’s overarching roadmap.

"Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence," reads Meta’s official corporate blog post introducing the models.

This substantial increase in scale and capability is also reflected in the commercial pricing structure for enterprise and developer access. Accessed through the Meta Model API, the standard Wonder model will carry a price tag of $4 per million input tokens, alongside $20 per million output tokens. This positions the model at a cost roughly equivalent to industry alternatives like GPT-6 Sol or Opus 5.5.

However, distribution for the more powerful Wonder Full model will follow a more restricted path. Notably, Wonder Full will only be made available to existing Muse Code subscribers and will not be served through the standard API. Furthermore, both newly released models will be integrated into Meta’s ecosystem to directly power the broader Muse agent platform.

True to Meta’s long-standing strategic interest in developing highly personalized virtual assistants capable of executing multi-step tasks over extended periods, the research team placed particular emphasis on a vital cognitive property they refer to as "active memory." This capability is designed to allow the models to reliably retain, recall, and synthesize information provided by the user across long spans of interaction, moving far beyond the short context windows typical of earlier AI generations.

To measure and validate these memory retention capabilities, Meta highlighted performance on the MemoryBench benchmark, a rigorous evaluation framework specifically designed to test continual learning and complex memory construction. According to the company’s internal testing data, Muse Wonder now achieves higher scores on MemoryBench than established competitors such as Fable, Opus, or Astra, indicating a substantial breakthrough in long-term contextual awareness.

"We put a tremendous amount of effort into making sure your model can remember what’s important to you over days, weeks, months, and beyond," explained Meta research scientist Kavya Kopparapu in an official video accompanying the Monday product launch.

The introduction of Muse Wonder and Muse Wonder Full marks the first true frontier model release since Scale AI founder Alexandr Wang stepped in to restructure and reshape Meta’s overarching artificial intelligence efforts in August of last year. Since that leadership and structural reorganization, Meta has steadily rolled out targeted components, including the lightweight Muse Spark model and the wider Muse agent ecosystem. Until now, however, the company had largely avoided the high-stakes, state-of-the-art model releases that have defined the public strategies of competitors like OpenAI and Anthropic, making Monday’s release a watershed moment for the company’s research division.

The computational foundation required to train these massive new systems came directly from Meta’s advanced infrastructure investments. According to the company’s announcements, both Wonder models were trained extensively at the Prometheus data center located in central Ohio, a massive facility boasting some of the highest operational capacities in the world that officially came online earlier this year.