Meta’s Superintelligence Lab has officially stepped into the frontier model arena with the release of a pair of advanced artificial intelligence systems, Muse Wonder and Muse Wonder Full. Announced on Monday, these new models represent a significant evolution for the company, moving beyond lightweight, low-latency applications to deliver heavier, highly capable systems designed to tackle complex computational challenges.
The launch follows the earlier introduction of Muse Spark, a model engineered primarily for speed and efficiency under low-latency constraints. By contrast, the Wonder models are substantially larger and exhibit superior performance metrics, boasting higher benchmark scores and notably enhanced capabilities in sophisticated tasks such as programming and code generation. According to testing data published in Meta’s official model card, both Wonder and Wonder Full achieve benchmark scores that are directly comparable to industry heavyweights like Fable 5.1 and GPT-6 Astra.
In a blog post accompanying the release, Meta characterized the development as a crucial milestone in its strategic roadmap. Together, the models mark the next step on the company’s scaling ladder and represent a major leap forward in the ongoing race toward achieving personal superintelligence.
This elevated scale is directly reflected in the commercial pricing structure established for the new release. Available for integration through the Meta Model API, the standard Wonder model is priced at $4 per million input tokens and $20 per million output tokens. This positions its operational cost on par with comparable frontier offerings such as GPT-6 Sol or Opus 5.5.
Distribution channels for the new technology vary by model tier. Notably, Muse Wonder Full will not be accessible through the standard API; instead, it will be restricted exclusively to subscribers of Muse Code. Furthermore, both Wonder and Wonder Full will serve as the underlying intelligence powering the broader Muse agent ecosystem.
In alignment with Meta’s long-standing strategic focus on developing autonomous personal assistants capable of executing multi-step tasks over extended periods, the company placed special emphasis on the models’ capacity to retain, recall, and synthesize user-provided information over time—a property that Meta has designated as "active memory."
To quantify and demonstrate these improvements in information retention, Meta highlighted performance on the MemoryBench benchmark, a rigorous evaluation framework designed to test continual learning and memory construction systems. According to the company’s internal testing results, Muse Wonder now surpasses Fable, Opus, and Astra on MemoryBench evaluations.
Elaborating on this capability in a launch video, research scientist Kavya Kopparapu noted the immense engineering effort dedicated to the project. The team poured a tremendous amount of work into ensuring that the model can reliably remember what is important to an individual user over days, weeks, months, and beyond.
The introduction of the Wonder models marks the first major frontier AI release since Scale AI founder Alexandr Wang undertook a substantial restructuring of Meta’s artificial intelligence divisions in August of last year. In the months following that reorganization, Meta introduced the lightweight Muse Spark model and rolled out features for the Muse agent, but deliberately steered clear of releasing the type of state-of-the-art frontier models historically associated with competitors like OpenAI and Anthropic.
The heavy computational requirements for training these sophisticated systems were met close to home. According to Meta’s announcement, Muse Wonder and Wonder Full were trained at the company’s Prometheus data center located in central Ohio, a massive facility featuring exceptionally high computing capacity that officially came online earlier this year.
Leave a Reply