Meta releases Muse Wonder, its first frontier AI model

Meta’s Superintelligence Lab released two new models on Monday — Muse Wonder and Muse Wonder Full — marking the company’s first attempt at a frontier AI model. Where the earlier Muse Spark model aimed for low-latency, lightweight performance, the Wonder models are larger and are said to be more capable, posting higher benchmark scores and better results on complex tasks such as coding. According to Meta’s own testing, presented in the model card, Wonder and Wonder Full achieve benchmark scores comparable to Anthropic’s Fable 5.1 and OpenAI’s GPT-6 Astra. “Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence,” Meta said in a blog post announcing the models.

Accessed through the Meta Model API, Wonder will cost $4 per million input tokens and $20 per million output tokens, putting it roughly in line with the cost of OpenAI’s GPT-6 Sol or Anthropic’s Opus 5.5. Wonder Full, the more capable of the pair, will not be served through the API at all — it will be available only to subscribers of Meta’s Muse Code service. Both models will also power Meta’s Muse agent.

In keeping with Meta’s push to build personal assistants capable of working over long stretches of time, the company placed particular emphasis on the new models’ ability to retain and synthesize information a user has shared with it, a property it calls “active memory.” To demonstrate that, Meta pointed to MemoryBench, a benchmark that tests continual-learning and memory-construction frameworks; the company said Muse Wonder now scores higher on MemoryBench than Fable, Opus, or Astra. Research scientist Kavya Kopparapu said in a video accompanying the launch that the team put significant effort into making sure the model could remember what matters to a user over days, weeks, and months.

Wonder and Wonder Full are Meta’s first frontier-model release since Scale AI founder Alexandr Wang restructured the company’s AI efforts in August of last year. Since then, Meta has shipped the lightweight Muse Spark model and its Muse agent, but had avoided the kind of state-of-the-art model releases for which OpenAI and Anthropic are known. According to the announcement, the new models were trained at Meta’s Prometheus data center in central Ohio, which came online earlier this year.

Why it matters

Meta’s first genuine attempt at a frontier model — and the memory-focused “active memory” push behind it — puts the company in direct competition with OpenAI and Anthropic on both raw capability and the long-horizon personal-assistant features likely to define the next phase of consumer AI products.