The launch follows months of structural and strategic changes within Meta’s artificial intelligence divisions. Where the earlier Muse Spark model was intentionally designed as a lightweight, low-latency alternative meant for quick tasks, the new Wonder models are substantially larger, significantly more capable, and built to handle highly complex computational challenges. According to benchmark testing released by the company alongside the models, both Wonder and Wonder Full demonstrate exceptional performance, achieving scores that place them in direct competition with industry-leading systems like Fable 5.1 and GPT-6 Astra.
"Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence," Meta’s official research blog post reads, highlighting the overarching ambition driving the development team in Ohio.
This newfound scale is immediately apparent in the commercial positioning and pricing structure of the models. Available through the Meta Model API, the standard Muse Wonder model will be priced at $4 per million input tokens and $20 per million output tokens. This places its cost roughly on par with other high-end enterprise offerings on the market, such as GPT-6 Sol or Opus 5.5.
However, Meta is taking a more restricted approach with its most powerful variant. Muse Wonder Full will not be served through the standard API; instead, access will be gated exclusively for Muse Code subscribers. Despite their differing access tiers, both models will play a foundational role in powering Meta’s broader ecosystem, serving as the underlying intelligence for the Muse agent.
In keeping with Meta’s long-term vision of developing sophisticated personal assistants capable of executing multi-step workflows over extended periods, the engineering team placed a heavy emphasis on a crucial capability: memory. Specifically, Meta is championing a property it calls "active memory," which refers to a model’s advanced ability to retain, recall, and synthesize information provided by users across extended interactions.
To validate these improvements, Meta pointed to the MemoryBench benchmark, a rigorous evaluation framework designed to test continual learning, long-term context retention, and complex memory construction. In the company’s internal testing, Muse Wonder managed to outperform competing models like Fable, Opus, and Astra on the MemoryBench metrics.
"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 research scientist Kavya Kopparapu in an accompanying video released for the launch.
The introduction of Muse Wonder and Muse Wonder Full marks the first major frontier model release since a significant leadership shuffle last year. Following a strategic restructuring of Meta’s AI efforts by Scale AI founder Alexandr Wang in August, the company largely focused on delivering lighter infrastructure, such as the Muse Spark model, and rolling out features for the Muse agent. Until now, Meta had largely stayed away from competing directly in the heavy-weight, state-of-the-art model releases that have defined the roadmaps of rival firms like OpenAI and Anthropic.
The massive computational requirements needed to train models of this scale also highlight Meta’s expanding physical infrastructure. According to the company’s announcement, Muse Wonder and Muse Wonder Full were trained at Meta’s Prometheus data center located in central Ohio, a state-of-the-art facility boasting immense capacity that officially came online earlier this year. As Meta continues to push the boundaries of model scale and active memory, the release of the Wonder models signals that the company is fully prepared to challenge the established leaders in the frontier AI landscape.
