Meta’s first frontier model, and the first one you cannot own. Muse Wonder 1.0 (October 12, preview) is the company’s “most capable model”: built for multi-file software engineering and long-horizon agentic work, with sub-agents, parallel tool calls, computer use, and every input kind Meta’s API takes. Context is 1,048,576 tokens. The tier below it in the family history, the Muse line, used to ship open weights; this one does not.
The split that matters. Muse Wonder runs on the pay-as-you-go Model API in Standard mode (Meta does not train on your content there). Muse Wonder Full, the stronger variant, exists only inside Muse Code subscriptions on the top usage plans, and setting reasoning effort to max on Wonder does not reach it. The frontier is a subscription; the API tier is a preview.
The open line stops at Glimmer. If you want Meta weights, Muse Glimmer 30B is still the open-weight answer; Wonder and Full are rented, at a reported 4 dollars and 20 dollars per million input and output tokens.