Meta Launches Frontier AI Models Muse Wonder, Targeting OpenAI and Anthropic

Meta's First Frontier Model Push

Meta's Superintelligence Lab unveiled two new large-scale AI models, Muse Wonder and Muse Wonder Full, marking the company's entry into the frontier model competition dominated by OpenAI and Anthropic. These models represent a significant step beyond Meta's previous lightweight Muse Spark offering, which prioritized low-latency performance. The Wonder models are substantially larger and built for more complex reasoning tasks, including advanced coding capabilities. According to Meta's internal benchmarking, the new models achieve performance levels comparable to OpenAI's GPT-6 Astra and Anthropic's Fable 5.1, two of the most advanced systems currently available.

Pricing and Distribution Strategy

Accessed through Meta's Model API, Muse Wonder carries a price tag of $4 per million input tokens and $20 per million output tokens, positioning it competitively with GPT-6 Sol and Opus 5.5 offerings. The company has adopted a tiered release strategy: Muse Wonder Full, the higher-performance variant, remains exclusive to Muse Code subscribers and will not be available through the standard API. Both models will power Meta's Muse agent, which the company is positioning as a personal assistant capable of executing tasks over extended timeframes. This distribution approach contrasts with OpenAI and Anthropic's typical broad API access models.

Memory as Core Innovation

Meta has placed particular strategic emphasis on what it calls "active memory"—the models' capacity to retain and synthesize information provided by users over extended periods. This capability reflects Meta's broader vision of building personal superintelligence assistants that can maintain coherent context across days, weeks, and months of interaction. The company developed and highlighted MemoryBench, a specialized benchmark designed to evaluate continual learning and memory construction frameworks. In testing, Muse Wonder outperforms competing systems from Anthropic and OpenAI on this metric, suggesting potential advantages for applications requiring long-term information retention and recall.

Context and Infrastructure

The Muse Wonder release marks Meta's first major frontier model announcement since August, when Scale AI founder Alexandr Wang restructured the company's AI operations. During this period, Meta released only the lightweight Muse Spark model and its agent framework, avoiding the competitive state-of-the-art model races that have defined recent AI industry dynamics. The new models were trained at Meta's Prometheus data center in central Ohio, which became operational earlier this year. This infrastructure represents Meta's substantial hardware investment in advanced AI development, positioning the company to compete more directly with OpenAI and Anthropic in the frontier model space.