Muse Wonder

Muse Wonder (released Oct 12, 2026) is the first frontier-class model from Meta's Superintelligence Labs, succeeding the lightweight Muse Spark line (Spark 1.3 scored 48 on the AA Intelligence Index six weeks earlier). It ships as two variants: Muse Wonder (model id muse-wonder-1.0, available via the Meta Model API and Muse Code) and Muse Wonder Full (stronger; Muse Code subscribers only — no API access or public price). The pitch is agentic work: "Active Memory" carries facts across conversations, and "Active Focus" keeps the model working through blockers and checking in before consequential actions. It accepts text, images, video, and audio. Unlike the Llama era, Wonder is closed — no weights release announced. It is positioned as the model that will power Meta's Muse personal agent (rolling out to subscribers first).

Provider-reported (Meta, model card): Meta's charts put Wonder Full ahead on several agent/coding benchmarks, e.g. 57.5% on Meta's SWE-Marathon (long-running software benchmark) vs 46.3% for Claude Opus 5.5; Meta also claims a lead on MemoryBench (continual-learning/recall) over Fable, Opus, and Astra.

Independent: Artificial Analysis Intelligence Index v4.3 scores Muse Wonder Full at 56 (second on the index, behind Claude Opus 5.5 at 58, ahead of GPT-6 Astra and Gemini 4 Argon at 53) and the standard Muse Wonder at 52; Full finished top-three on six of the ten index evaluations. AA pegs cost at $4.52 per benchmark task (vs $1.60 for Muse Spark 1.3). Independent hands-on coverage is still thin — the model is one day old — and Meta's safety report is promised but not yet published (vandatateam).