Muse Wonder

Muse Wonder, released October 12, 2026 by Meta Superintelligence Labs alongside Muse Wonder Full, is Meta's first frontier-class model. The API model ID is muse-wonder-1.0, served on the Meta Model API with a 1,048,576-token context window; inputs are text, images, video, and audio (PDFs via upload), output is text.

Meta markets two trained behaviors. Active Memory keeps user context across long conversations and surfaces the details that matter for the current task (Meta's internal MemoryBench scores both Wonder models ~85.5, ahead of every model Meta tested). Active Focus keeps the model on a goal — trying other approaches, working through errors, asking for clarification. Reasoning effort has six settings from minimal to max, and the API speaks OpenAI Responses, OpenAI Chat Completions, and Anthropic Messages formats.

Pricing is $4 per million input tokens, $0.20 cached, $20 output — exactly Anthropic's Claude Opus 5.5 rate card. Meta's own 14-row benchmark table is unusually well documented but mixes independent results with benchmarks Meta built and graded itself: the Wonder family leads 7 rows, Opus 5.5 leads 6, GPT-6 Astra 1, and the standard Wonder beats Opus 5.5 on 5 of 14. The independent read is Artificial Analysis's Intelligence Index: Wonder 52, Wonder Full 56, Opus 5.5 58.

Provider-reported (Meta, 14-row table): Wonder family leads 7 rows, Claude Opus 5.5 leads 6, GPT-6 Astra 1; standard Wonder beats Opus 5.5 on 5 of 14. Independent (Artificial Analysis Intelligence Index): Wonder 52, Wonder Full 56, Opus 5.5 58.