Muse Wonder: Meta Won Every Test It Made Itself

Meta says Muse Wonder puts it back at the AI frontier. We read the 13-page model card and checked all 14 benchmarks Meta picked. On the 2 tests Meta built itself, Muse wins 2 of 2. On the 12 built by others, it wins 5. QUICK FACTS (checked Oct 12, 2026) Two models: Muse Wonder (Meta Model API + Muse Code, $4 / $20 per million tokens) and Muse Wonder Full (Muse Code Power $50/mo or Maximum $200/mo only). 1M-token context; reads text, images, video and audio. Closed weights, hosted by Meta only. Meta's table: Muse leads 7 of 14 rows; Claude Opus 5.5 leads 6. No Gemini model is compared. Independent: Artificial Analysis Intelligence Index puts Wonder Full at 56 (second), behind Claude Opus 5.5 at 58. Muse Wonder scores 52; Gemini 4 Argon 53. Vals AI Finance Agent V2 leaderboard: Gemini 4 Argon 65.40%, Muse Spark 1.2 60.60%; Meta reports 59.1% for Wonder Full (different runs). Model card: ExploitGym 92 flags, 21 through the planted bug, 71 through unintended holes. Without mitigations Meta can't rule out its high-risk threshold for chem/bio. Full safety report: "available soon." Muse Code subscriptions default to Contributor mode (Meta may train on your content). OUR TEST We sorted Meta's own 14 benchmark rows by who made each test (from Meta's model card) and compared the best Muse score with the best rival. Real output: Meta-made tests, Muse wins 2 of 2; outside-made tests, 5 of 12. Code shown on screen. CHAPTERS 0:00 Meta says it's back at the frontier 0:55 What Meta shipped: Wonder vs Wonder Full 2:04 How Meta got here 3:29 Active Memory + Active Focus 5:02 The 14 benchmarks + our own check 7:41 What independent testers say 8:31 The safety report: 92 flags, 71 surprises 9:51 Price, plans + the contributor-mode catch 10:48 Claim vs reality + my take 11:41 Your turn SOURCES (everything shown on screen) Meta newsroom: https://about.fb.com/news/2026/10/int... Meta research blog: https://research.meta.ai/blog/muse-wo... Muse Wonder model card (PDF): https://research.meta.ai/static/muse-... Meta developer page (table, prices, plans): https://dev.meta.ai/models/muse-wonder Meta docs, models: https://dev.meta.ai/docs/models Muse Code subscriptions: https://dev.meta.ai/docs/muse-code/su... Vals AI Finance Agent V2: https://www.vals.ai/benchmarks/fabv2 Artificial Analysis, Wonder Full vs Opus 5.5: https://artificialanalysis.ai/models/... Artificial Analysis, Muse Wonder: https://artificialanalysis.ai/models/... TechCrunch, launch: https://techcrunch.com/2026/10/12/met... CNBC: https://www.cnbc.com/2026/10/12/meta-... TechCrunch, Llama 4 benchmarks (2025): https://techcrunch.com/2025/04/06/met... TechCrunch, Muse app downloads: https://techcrunch.com/2026/09/25/met... Observer, Wang's "appetizer" interview: https://observer.com/2026/06/alexandr... Mark Zuckerberg on X: https://x.com/finkd/status/2109689305... · https://x.com/finkd/status/2109689307... AI at Meta on X: https://x.com/AIatMeta/status/2109691... Alexandr Wang on X (Sphere video): https://x.com/alexandr_wang/status/21... Meta videos:    • Introducing Muse Wonder & Muse Wonder Full...   ·    • Muse Wonder and Muse Wonder Full: New Mode...   ·    • Take the full tour of Muse, Meta's persona...   Benchmark numbers are the labs' own unless marked independent. Would you let Meta's agent run your email for $20 a month? Tell me in the comments. FOOTAGE: Meta's own videos, demos and posts (shown for commentary and fact-checking); page footage recorded from each source; illustration scenes generated by AI Decoded. MUSIC: "Voxel Revolution" and "Digital Lemonade" by Kevin MacLeod (incompetech.com), licensed under CC BY 4.0. Narration uses an AI-generated voice. #AI #Meta #MuseWonder