Late on 12 October, Meta introduced Muse Wonder and Muse Wonder Full. They are built for agentic work: longer tasks where AI plans several steps, uses tools and tries to recover when something goes wrong.
Meta says the models match leading alternatives in areas including coding, memory and professional tasks. Most of those comparisons come from the company itself, so their practical value still needs independent testing.
The standard Muse Wonder is now available in Muse Code and through the Meta Model API at $4 per million input tokens and $20 per million output tokens. The more capable Full version requires a Power or Maximum Muse Code subscription and is not available through the API.
Neither model has open weights that you can download and run locally. Meta is therefore separating smaller, more open models such as Muse Glimmer from its strongest cloud models. Muse Wonder will later power the personal Muse assistant, starting with subscribers, but Meta has not given a date.
That is the important detail behind the launch language: Meta is back in the frontier race, but not with a new open Llama successor.