Meta releases Muse Wonder and Muse Wonder Full AI models
Meta has launched its first frontier AI models, Muse Wonder and Muse Wonder Full, aiming to compete with top models from OpenAI and Anthropic.
Meta’s Superintelligence Lab released two new state-of-the-art AI models on Monday named Muse Wonder and Muse Wonder Full. Announced by the company, these releases represent Meta's first attempt at a frontier model following a major restructuring of its artificial intelligence efforts led by Alexandr Wang and CEO Mark Zuckerberg.
The newly released models are larger and more capable than Meta's previous lightweight offering, Muse Spark. According to Meta's testing, the Wonder models achieve high benchmark scores and excel at complex tasks like coding and working with legal documents.
What are Muse Wonder and Muse Wonder Full?
Muse Wonder and Muse Wonder Full are designed to rival the industry's best artificial intelligence models from competitors like OpenAI and Anthropic. Meta states that its internal testing shows the new models achieve comparable or better benchmark scores than models such as GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5.
The models were trained at Meta’s Prometheus data center in central Ohio, which came online earlier this year. They are designed to support personal AI assistants and will soon be used to power the Muse agent.
How can businesses access the new models?
Access to the new models varies depending on the specific version and subscription type. Pricing and availability details include the following options for developers and users:
Muse Wonder is available through the Meta Model API at a cost of four dollars per million input tokens and twenty dollars per million output tokens.
Muse Wonder Full is exclusively available to Muse Code subscribers and is not served through the API.
Both models will eventually power features across Meta's platforms, starting with subscribers to the Muse agent.
Why active memory matters for businesses
Meta placed particular emphasis on the models' ability to retain and synthesize user information over extended periods, a property the company calls active memory. Alongside a companion feature called active focus, these tools are built to help personal assistants perform complex tasks over days, weeks, or months without losing track of their goals.
To measure this capability, Meta highlighted the MemoryBench benchmark, which challenges continual learning frameworks. According to company testing, Muse Wonder scores higher on MemoryBench than competing models from Fable, Opus, and Astra, potentially reducing expensive mistakes or compounded errors during complex research and software projects.
Comments 0
No comments yet. Start the conversation.