Muse Wonder

Meta Superintelligence Labs' agent-first frontier line, launched as two proprietary variants on one model card: Muse Wonder (muse-wonder-1.0, Meta Model API, $4.00 / $20.00 per 1M input/output and $0.20 cached input) and Muse Wonder Full (muse-wonder-full-1.0, subscription-only through Muse Code). Both take text, image, video and audio in and emit text, over a 1M-token context. Meta pitches two named behaviours rather than a size: Active Memory, "carrying important context about you across conversations and actively retrieving the details relevant to the task at hand," and Active Focus, "continuing working toward a challenging goal: trying alternative approaches, working through blockers, and adapting to new information."

The launch scorecard compares both variants at max reasoning effort against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5, and it is not a sweep. Wonder Full leads the group on professional agent work and long-horizon engineering: Harvey LAB-AA 13.1 against Astra's 8.6, Fable 5.1's 6.4 and Opus 5.5's 4.2 — half again the best rival on legal deliverables, and more than triple Opus — plus SWE-Marathon 57.5 (xhigh) against 42.5–46.3, DeepSWE 1.1 79.0, AutomationBench-AA 45.4, and Meta's internal Personal Superintelligence Index 59.1. On MemoryBench, the evaluation built for the Active Memory claim, the smaller Wonder edges Full 85.6 to 85.5, both clear of the 79.0–81.5 field.

It trails Opus 5.5 where the comparison is least favourable to the pitch. Meta's own AA Intelligence Index row (v4.3.2) reads 56 for Wonder Full and 52 for Wonder, against 58 for Opus 5.5; ProgramBench is 59.0 vs 70.5, OSWorld V2.1 78.4 vs 81.8 partial, Terminal-Bench 4.0-AA 57.1 vs 59.6, and HealthBench Pro 62.3 vs 65.6. Vals Finance Agent V2 splits the family: Full takes it at 59.1 over Fable 5.1's 58.9 and Opus 5.5's 58.6, while the base Wonder's 56.6 trails both. So the frontier claim holds for agentic and legal work and for sustained coding, not for general intelligence. Artificial Analysis published its own pages the day after launch and confirmed Meta's numbers exactly on AA Intelligence Index v4.3.2: 52 for Muse Wonder (max reasoning), ranking 9th of 216 scored models, and 56 for Muse Wonder Full, ranking 4th — Meta's best placement to date and a jump from Muse Spark 1.3's 48 on that version. On v5.0 (October 27) they read 43 for Muse Wonder and 50 for Muse Wonder Full, against 58 for Opus 5.5. On v5.0.1, after November's CritPt v2 regrade, Muse Wonder reads 44 and Full stays at 50, against Opus 5.5's 59; AA names Full the leader on GDP.pdf (33% all-pass). AA lists Wonder at $4.00 / $20.00 per 1M and notes Full has no publicly available API, consistent with its subscription-only release.

Released under the Meta Superintelligence Scaling Framework, with an unusually direct disclosure: Meta "does not rule out the model reaching" the high-risk threshold in the Chemical & Biological domain without mitigations, and so ships safety post-training plus system-level classifiers for that domain; Cybersecurity and Loss of Control are assessed "moderate or lower" even unmitigated. Cyber capability on Wonder Full: CyberGym 88.4%, ExploitBench 58.9%, and 92 flags on ExploitGym — of which a verifier pass found only 21 used the intended vulnerability, the figure comparable to the published ranking. Parameters are undisclosed and no weights were released, though Mark Zuckerberg has said Meta plans open-weights versions of the Muse line.