Meta has launched the Muse Wonder series, introducing Muse Wonder and Muse Wonder Full as its most capable artificial intelligence models to date. The release expands on the earlier Muse Spark model and moves the company into the frontier model category dominated by OpenAI and Anthropic. Meta says the new systems deliver stronger coding, reasoning, and complex task performance along with long-term memory features for more personalized assistants.
Advanced Performance for Frontier Tasks
The two models represent Meta's first entry into the frontier model category, which combines high computational capability with strong results on difficult assignments. Muse Wonder and Muse Wonder Full are designed to handle expert domains such as law, finance, and health, while also supporting developers who are building applications or games. Internal company tests indicate that performance is close to competing systems such as Fable 5.1 and GPT-6 Astra, although full comparative benchmarks were not disclosed.
Active Memory and Focus
A central feature of the new series is Active Memory, which allows the models to retain important user information across conversations for days, weeks, or months. Meta stated that this capability supports long-term projects by recalling preferences and relevant details instead of relying only on the current chat context. The company also introduced Active Focus, which helps the models keep working toward goals, recover from errors, and ask for clarification before taking important actions.
Agentic Capabilities and Personalization
The models are designed to act as personal agents that can work across messages, calendars, and documents to complete real tasks. Meta said Muse Wonder and Muse Wonder Full can interpret complex information such as insurance claims, leases, and personal finances before working through each problem. This agentic direction reflects the company's broader goal of moving beyond simple question answering toward assistants that manage ongoing work.
Pricing and Access
Muse Wonder is available through the Meta Model API and Muse Code, with API pricing set at US$4 per million input tokens and US$20 per million output tokens. Muse Wonder Full will not be offered through the API and will be limited to subscribers of the Muse Code service. Meta plans to use both models to power its Muse Agent assistant and will roll out enhanced capabilities to subscribers first.
MemoryBench and Safety
Meta evaluated the models using MemoryBench, a benchmark that measures continuous learning, memory construction, and information retrieval. According to the company, Muse Wonder outperformed Fable, Opus, and Astra on this memory-focused test, reflecting the emphasis on persistent personalization. For safety, the models were trained to avoid sharing private information without permission, decline requests that could cause real harm, and strengthen defenses against cyber attacks.
Strategic Context
The launch follows a restructuring of Meta's AI efforts led by Alexander Wang, founder of Scale AI, in August 2025. Since then, Meta has released Muse Spark and the Muse Agent assistant before advancing to the more capable Muse Wonder models. The new systems were trained in the Prometheus data center in Ohio, which began operating this year and highlights the growing importance of computing infrastructure.
Meta positions the Muse Wonder series as a major step toward what it calls personal superintelligence, where assistants can understand user needs and act as digital agents over time. By combining frontier performance with persistent memory and safety controls, the company is targeting both individual users and developers who need reliable AI support for complex work. The models are available now through Meta's developer channels, with broader availability expected across Meta services in the coming period.