Muse Wonder

Muse Wonder, released October 12, 2026, opened the Muse Wonder series, a class of models Mark Zuckerberg described as much larger and more advanced than Muse Spark and the one Meta said brought it to the frontier. It launched beside Muse Wonder Full, as the faster of the two, fourteen months after Meta began rebuilding its AI effort from the ground up. Meta trained both at Prometheus, its gigawatt-scale data center, and built the launch around two behaviours it named itself: Active Memory, carrying context about a user across conversations and pulling it back when it becomes relevant, and Active Focus, staying on a long task through blockers and checking in before consequential actions. It was the half of the pair developers could reach, available on day one through the Meta Model API and Muse Code with a 1M-token context window and text, image, video, audio and PDF input, while Meta said Muse itself would switch to it later, subscribers first.

Meta's launch table ran it at max reasoning against Muse Wonder Full, Claude Opus 5.5, Claude Fable 5.1 and GPT-6 Astra. Its best result was 85.6% on MemoryBench, the top score on the table and a tenth of a point ahead of Muse Wonder Full, and it scored 78.8% on DeepSWE 1.1 against 74.2% for Opus 5.5 and 10.0% on Harvey's Legal Agent Benchmark as run by Artificial Analysis, ahead of all three rival models. Elsewhere the smaller model gave ground: it trailed both Anthropic models on JobBench at 70.5% and Finance Agent v2 at 56.6%, and Opus 5.5 and GPT-6 Astra on AutomationBench at 39.1%. Meta put its Artificial Analysis Intelligence Index at 52, against 56 for Muse Wonder Full and 58 for Opus 5.5.