Muse Wonder

Not yet independently verified. Every detail here comes from Meta's own announcement of 12 October 2026 and its Model API documentation; no independent report was found. The announcement names Harvey's Legal Agent Benchmark, FocusBench and SWE-Marathon but gives no figures, and the model is described as a preview, so no score is stored. Weights are not published and no licence is stated. We will update this when it can be confirmed, and remove this note.

Muse Wonder (model ID muse-wonder-1.0) was announced by Meta on 12 October 2026 as its most capable model for personal and professional agentic tasks. Meta's documentation describes it as a preview aimed at complex software engineering and long-horizon agent work. It accepts text, images, video, audio and PDF files, returns text, and has a context window of 1,048,576 tokens. Reasoning effort can be set from minimal to max. It is available through the Meta Model API on pay-as-you-go in standard mode only, and in Muse Code on every subscription plan; Meta says it will soon power experiences in Muse, starting with subscribers. The API price is $4.00 per million input tokens, $20.00 per million output tokens and $0.20 per million cached input tokens, with no long-context premium and no peak-hours rate. A larger variant, Muse Wonder Full, is available only in Muse Code on the Power Usage and Maximum Usage plans and is not offered through the API; setting reasoning effort to max on Muse Wonder does not select it. Meta's announcement names Harvey's Legal Agent Benchmark, FocusBench and SWE-Marathon and claims leading results on them, but gives no figures in the post, so none are stored here. Meta also says the model was trained at its Prometheus data centre and promises a safety report soon. Weights are not published and the documentation states no licence.