Meta’s Superintelligence Lab introduced Muse Wonder and Muse Wonder Full, their first frontier AI models designed for higher capability and better retention of user information, targeting advanced coding and continuous learning tasks.
- Muse Wonder models excel in complex tasks such as coding and continual learning.
- Active memory feature supports long-term retention and synthesis of user data.
- Wonder priced at $4/$20 per million tokens; Wonder Full exclusive to Muse Code subscribers.
What happened
Meta’s Superintelligence Lab released two new frontier AI models named Muse Wonder and Muse Wonder Full. These models represent a significant upgrade from the company’s previous lightweight Muse Spark model, offering enhanced capabilities, including superior performance in complex tasks like coding. Benchmark results show that the Wonder models perform comparably to other leading models such as Fable 5.1 and GPT-6 Astra.
The new models were trained at Meta’s Prometheus data center in Ohio, marking the company’s first frontier model release since a restructuring led by Scale AI founder Alexandr Wang last year. Muse Wonder is available through Meta’s Model API with specific pricing, while Muse Wonder Full is limited to Muse Code subscribers and not available via the API. Both variants will be integrated into Meta’s Muse agent platform.
Why it matters
Meta’s introduction of Muse Wonder models signals its renewed commitment to advancing large-scale AI capabilities alongside industry leaders. Unlike the previously released Muse Spark, which emphasized efficiency and low latency, Wonder focuses on scaling power and improved task sophistication, particularly in coding and memory retention. This helps position Meta competitively in an increasingly crowded AI landscape.
A standout feature of these models is their enhanced 'active memory' capability, allowing them to remember and synthesize important user data for extended periods, spanning days to months. The models outperform competitors on Meta’s MemoryBench benchmark, indicating a breakthrough in continual learning—a key feature for personal AI assistants designed to evolve with user interactions.
What to watch next
Market attention will focus on how widely Meta’s Muse Wonder models are adopted, especially given the tiered access and pricing structure. The exclusivity of Wonder Full to Muse Code subscribers suggests Meta aims to engage developers and businesses interested in advanced coding tasks and personal assistant applications. Usage and feedback from this cohort will be critical in shaping future releases.
Additionally, the integration of these models into the Muse agent platform will provide insights into their real-world effectiveness in personal AI tasks requiring long-term memory retention and complex problem-solving. Observers should also track Meta’s ongoing AI strategy following recent organizational changes and how its frontier models compete against industry giants like OpenAI and Anthropic.