Meta on Monday, Oct. 12, released Muse Wonder and a stronger version, Muse Wonder Full, which it calls its most capable AI models. Meta says both lead rival models at agentic tasks for people's personal and working lives, meaning tasks the AI carries out on its own, from planning a run to writing software. Meta says that 14 months ago it began rebuilding its AI work from the ground up, and that it is now pushing the frontier, meaning it builds AI models that match or beat the best that exist today.
Developers can use Muse Wonder now through Meta's API, a connection that lets developers plug Meta's models into their own apps. It costs $4 for every million tokens it reads and $20 for every million it writes, tokens being the small chunks of text, often parts of words, that AI models read and write. Meta lists the model as a preview. It is also in Muse Code, Meta's coding agent, while Muse Wonder Full is only in Muse Code, on the $50 and $200 monthly plans. Muse Code subscriptions are set by default to let Meta use people's prompts and code to train future models, in exchange for more use, and subscribers can turn that off. Meta says Muse Wonder will soon power features in Muse, Meta's AI assistant, starting with subscribers, and that it will later use Muse Wonder to build specialized models for everyone who uses Muse. It has not given dates.
Meta built the models around two skills it calls Active Memory and Active Focus. Meta says the first carries important details about a person from one conversation to the next, and the second keeps the model working toward a hard goal, trying other approaches when it gets stuck. Meta says the model asks for clarification when needed and checks in before taking important actions. Meta's demos show Muse Wonder building a running plan around a person's preferences, laying out a wedding seating chart and building a playable 3D kitchen game. Meta trained the models at Prometheus, a data center it describes as gigawatt-scale.
Meta's comparison pits its models against OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 and Claude Opus 5.5, mixing scores from outside testing groups and the rival companies' own reports with its own runs, on a set that includes several tests Meta built itself. On Meta's own results, only Muse Wonder Full comes out ahead on its overall score for tasks done for people; the standard Muse Wonder lands just behind the best OpenAI and Anthropic models there, and rivals still beat Full on some tests, such as answering health questions and using a computer. Where Meta ran rival models itself, it says it may not have set them up as well as possible. On one independent overall ranking of nearly 700 AI models, Muse Wonder Full places fourth. Meta says the models can still slip up on tasks with many steps, especially unfamiliar ones, that judging space and shapes, especially in 3D, is a known weak spot, and that they are not meant for anyone under 18.
Meta's model card, its published overview of how the model was built and tested, says that without safeguards Meta could not rule out its strongest version reaching the company's high-risk level for chemical and biological threats, the point at which releasing it could substantially help bring about a threat Meta considers severe. Meta says extra safety training and automated filters bring that risk down to moderate or lower, which it says supports releasing the model. Meta says a fuller safety report is coming soon.