Meta debuts its first frontier model, Muse Wonder.

Meta description: Meta Superintelligence Lab released Muse Wonder and Muse Wonder Full, its first frontier models, with $4 per million input token API pricing on Monday.

Tags: Meta Superintelligence Lab, Muse Wonder, Muse Wonder Full, active memory, MemoryBench

Featured image: Abstract data center and memory network visualization, no real people, no company logos, no text.

Meta Superintelligence Lab released Muse Wonder and Muse Wonder Full on Monday, marking the company’s first attempt at a frontier model. The pair moves Meta beyond the lightweight Muse Spark line toward larger systems that are presented as more capable on complex tasks, including coding, and that are compared in benchmark testing with Fable 5.1 and GPT-6 Astra. Their API pricing is $4 per million input tokens and $20 per million output tokens, and their design emphasizes “active memory” for assistants that operate over long periods.

Background: Meta Rebuilt Its AI Effort Before the Frontier Push

Before the Monday launch, Meta’s AI work had been moving through a restructuring that began in August of last year, when Scale AI founder Alexandr Wang reorganized the company’s AI efforts. After that change, Meta introduced Muse Spark, a model built to be lightweight and to operate at low latency, and the Muse agent. It did not, however, enter the class of state-of-the-art releases associated with OpenAI and Anthropic.

The company’s interest in personal assistants that can perform tasks over a long period made memory a more visible design goal. Such assistants must not only answer quickly; they must retain and synthesize information given by the user across later interactions. Meta calls that property “active memory.”

Development: Muse Wonder and Muse Wonder Full Target Higher Benchmarks

The new models, Muse Wonder and Muse Wonder Full, are described as larger than Muse Spark and more capable, with higher benchmark scores and better results on complex tasks such as coding. In testing presented in the model card, Wonder and Wonder Full achieve comparable benchmark scores to Fable 5.1 and GPT-6 Astra.

Meta’s blog post placed the launch within a staged scaling strategy: “Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence.”

Memory is the other central claim. Meta pointed to MemoryBench, a benchmark that challenges continual learning and memory construction frameworks. According to the company’s testing, Muse Wonder now scores higher on MemoryBench than Fable, Opus, or Astra. Research scientist Kavya Kopparapu said in a video accompanying the launch, “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.”

Latest Progress: API Pricing and Training Infrastructure

The scale of the models is reflected in their pricing. Accessed through the Meta Model API, Wonder will cost $4 per million input tokens and $20 per million output tokens, making it roughly equivalent in cost to GPT-6 Sol or Opus 5.5. For users evaluating the release, the practical comparison points are therefore benchmark results, memory performance on MemoryBench, and the cost per million tokens.

The models were trained at Meta’s Prometheus data center in central Ohio, which came online earlier this year. That detail ties the frontier model release to the company’s AI infrastructure timeline: the models appeared after a data center that had become operational in the same year.

Current Stage: A First Frontier Model Release After the Restructuring

Muse Wonder and Muse Wonder Full represent the first frontier model release since the August restructuring, following the earlier Muse Spark model and Muse agent. The launch gives Meta a larger, benchmark-focused pair of models, an active-memory story tied to MemoryBench, API pricing of $4 per million input tokens and $20 per million output tokens, and a training location in the Prometheus data center.

What to watch now is whether the models’ performance on complex tasks and their memory claims hold up in use, and whether the pricing places the Wonder models in practical competition with GPT-6 Sol and Opus 5.5 in the frontier-model landscape.