Meta introduced Muse Wonder and Muse Wonder Full on October 12, calling them its most capable models for personal and professional agent work. The company says they are intended for tasks such as coding, analyzing financial or legal material, and carrying work across multiple steps.
Meta describes two features as central to that pitch. Active Memory is meant to retain relevant details across conversations, while Active Focus is designed to keep an agent working toward a goal, try another approach after a failure, and ask before taking consequential actions. Those are company claims about how the models behave, not guarantees that long-running tasks will succeed.
In one third-party benchmark suite, both models improved on Meta's previous generation. Artificial Analysis gave Wonder and Wonder Full scores of 52 and 56, respectively, on version 4.3 of its Intelligence Index at maximum reasoning effort. Muse Spark 1.3 scored 48 under the same setting. The largest gains appeared in agent-focused evaluations, though one benchmark suite cannot establish performance across every real-world workflow.
The models accept text, images, audio, and video and return text, with context windows of about 1 million tokens, according to VentureBeat. Neither model is open weight, a departure from the downloadable approach associated with some earlier Meta releases.
Muse Wonder is available through the Meta Model API and Muse Code. API pricing is $4 per million input tokens, $20 per million output tokens, and $0.20 per million cached input tokens. Muse Wonder Full, the higher-scoring version in Artificial Analysis' tests, is limited to Muse Code subscribers and does not have public pay-as-you-go API access.
Meta says Muse Wonder will later power experiences in its Muse assistant, beginning with subscribers. It also says the models include safeguards for private information, cyber misuse, and harmful requests, while noting that a fuller safety report is still to come. Whether the benchmark gains translate into dependable agent work outside Meta's demonstrations and controlled evaluations remains unresolved.
