Muse Wonder is a model Meta released on October 12, 2026, alongside a higher-tier version called Wonder Full, built for long-running agent tasks. Its two signature capabilities are Active Memory and Active Focus: the first carries important things across conversations, the second tries other approaches when blocked and sticks with it to the end. The standard version is billed per use through the Meta Model API, Wonder Full only comes inside Muse Code subscriptions, and the free version does not get it.
What Active Memory and Active Focus actually are
Active Memory brings important context back across conversations: your preferences and the details of tasks in progress, so you do not have to explain them again next time. Active Focus means not giving up directly when a task stalls: if a tool call fails, try another approach; if a page structure changes, adjust the plan; if it gets stuck, ask you; and before important actions, check with you first. These two capabilities are built for long tasks; in short conversations you will barely notice a difference.
Prices and access: look before you leap
The context window is 1 million tokens, input supports text, image, video, and audio, and output is text. API pricing is $4 per million input tokens and $20 per million output tokens, and the model ID is muse-wonder-1.0. Wonder Full has no standalone API price; it only comes in the Muse Code Power Usage ($50 a month) and Maximum Usage ($200 a month) subscriptions. Ordinary users will have to wait for it to arrive in the Muse app, with paid subscribers first; South Korea is currently not covered.
Do not be led astray by launch claims
Meta has disclosed neither the parameter count, the training data, nor the weights, and has provided no independent benchmark scores. MemoryBench and FocusBench are both Meta's own internal benchmarks, so the claim of being "on par with the frontier" is the vendor's statement, not an independent conclusion. How memories are stored, how long they last, and how they are deleted is not fully explained in the public materials, so do not put sensitive information into it. Developers who want a taste should go through the API and start with small volumes; ordinary users should wait until it lands in the Muse app and judge its real-world performance then.
