Meta’s Superintelligence Lab has made a splash with the release of two new frontier models, Muse Wonder and Muse Wonder Full. These models are considered to be larger and more capable than their predecessor, Muse Spark, offering better results on complex tasks like coding.

The Wonder models, with benchmark scores comparable to Fable 5.1 and GPT-6 Astra, are part of Meta’s scaling ladder towards personal superintelligence. However, the pricing is a step up, with Wonder costing $4 per million input tokens and $20 per million output tokens. Muse Wonder Full is exclusively for Muse Code subscribers.

Meta’s focus on retaining and synthesizing user information, known as “active memory,” is highlighted. The company’s use of the MemoryBench benchmark shows that Muse Wonder now scores higher than Fable, Opus, or Astra in continual learning and memory construction frameworks.

The new models mark a significant shift since Scale AI founder Alexandr Wang restructured Meta’s AI efforts in August last year. Since then, Meta has released Muse Spark and the Muse agent, but avoided the kind of state-of-the-art model releases we’ve seen from OpenAI and Anthropic.

The training of these models took place at Meta’s Prometheus data center in central Ohio, which has been operational since earlier this year.