Meta announced the release of two new models, Muse Wonder and Muse Wonder Full, marking the company’s first attempt at a frontier model. The launch, which occurred Monday, introduces models described as significantly more capable than Meta’s earlier, lightweight Muse Spark.
These Wonder models are larger and show higher benchmark scores. They perform better on complex tasks, particularly coding.
Meta claims Wonder and Wonder Full achieve benchmark scores comparable to Fable 5.1 and GPT-6 Astra, positioning them among leading models available today.
Beyond general capability, Meta placed emphasis on the models’ ability to handle sustained interaction. This functionality, which the company calls “active memory,” involves retaining and synthesizing information provided by users over extended time periods.
To measure this feature, Meta highlighted the MemoryBench benchmark. According to the company’s own testing, Muse Wonder achieved higher scores on MemoryBench compared to Fable, Opus, or Astra.
Research scientist Kavya Kopparapu explained that Meta dedicated substantial effort to ensuring the model could remember important user details over months and weeks.
The models will power Meta’s Muse agent, but their availability and cost differ significantly between the two variants. Muse Wonder will be accessible through the Meta Model API. Using this API, Muse Wonder costs $4 per million input tokens and $20 per million output tokens.
This pricing makes the standard Wonder model roughly equivalent in cost to GPT-6 Sol or Anthropic’s Opus 5.5. The higher-powered Muse Wonder Full variant will not be served through the API. Instead, it will be available exclusively to Muse Code subscribers, positioning it as a premium offering.
The launch signals a strategic shift for Meta. Since restructuring its AI efforts in August of last year, the company had focused on the lighter Muse Spark model and the Muse agent. The Wonder series marks a departure toward directly competing at the frontier of AI model development.
The Wonder models were trained at Meta’s Prometheus data center, a facility located in central Ohio that commenced operations earlier this year.
