Skip to content
Türkçe

Meta · Models · Agents

Meta unveils Muse Wonder and Muse Wonder Full, its most capable models yet

Published: 22 sourcesTürkçe

Meta unveils Muse Wonder and Muse Wonder Full, its most capable models yet

On October 12, Meta announced the first frontier-class models from Meta Superintelligence Labs. They’re built for AI agents that handle personal and professional work on people’s behalf. Meta highlights two features: Active Memory lets the model remember what matters for weeks, and Active Focus makes it try another route instead of giving up when it hits a blocker.

The stronger Muse Wonder Full isn’t open to developers

Muse Wonder is in preview on Meta’s API at $4 per million input tokens and $20 per million output tokens, the same list price as Claude Opus 5.5. Wonder Full isn’t on the API at all; it’s only in the $50 and $200 monthly plans of Muse Code, Meta’s coding agent. Both models have closed weights.

In Meta’s table, Wonder Full leads on agent work

In Meta’s own comparison, Wonder Full scored 13.1 on Harvey LAB-AA, a legal agent test, more than three times Claude Opus 5.5’s 4.2. It also leads on AutomationBench, a business workflow test. The Personal Superintelligence Index at the top of the list, though, is Meta’s own internal test.

Claude Opus 5.5 leads on six tests

Muse models lead 7 of the table’s 14 rows, Claude Opus 5.5 leads 6 and GPT-6 Astra one. Coding is split: Wonder Full wins DeepSWE and SWE-Marathon, while Opus wins Terminal-Bench and ProgramBench and leads the Artificial Analysis index 58 to 56. Claude Sonnet 5.5 and Gemini 4 Argon aren’t in the comparison.

Independent rankings put it in the top five too

On Artificial Analysis’ index, Wonder Full scores 56 and ranks 4th of 690 models, behind two versions of Claude Opus 5.5 and Claude Sonnet 5.5. It’s 5th of 48 on Vals AI’s index, where Vals said Meta now tops every OpenAI model. On LMArena’s text leaderboard, Muse Wonder is 2nd, though with only 1,944 votes so far.

Zuckerberg describes Meta’s $100-200 billion AI bet

On Stripe co-founder John Collison’s Cheeky Pint podcast, Zuckerberg said that at the end of 2025, before seeing the models, Meta decided to “write this check for like, I don’t know, like $100 billion, $200 billion.” It isn’t a payment but spending on data centers and computing power; Meta’s 2026 capital spending plan is $130-145 billion.

But the strongest version is closed to developers

Because Wonder Full isn’t on the API, developers can’t build it into their own products, and Artificial Analysis lists it as having no publicly available API. Meta’s detailed safety report isn’t out yet either. Some users have complained that the Muse agent accessed their messages without permission, which Meta denies.

Sources

  1. Meta Newsroom, “Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents”, Official announcement; availability, Active Memory and Active Focus, safety, (about.fb.com)
  2. Meta AI Research, “Muse Wonder: A Frontier Model For Personal & Professional AI Agents”, Technical blog; benchmark table and charts, videos, (research.meta.ai)
  3. Meta AI Research, “Muse Wonder Model Card”, Model card: 1M context, methodology, SWE-Marathon note, risk assessment, (research.meta.ai)
  4. Meta Model API Docs, “Models - Meta Model API”, muse-wonder-1.0 in preview; Wonder Full only in Muse Code (dev.meta.ai)
  5. Meta Model API Docs, “Pricing and rate limits”, $4 input and $20 output per million tokens (dev.meta.ai)
  6. Meta Model API Docs, “Subscriptions”, Muse Code plans; Wonder Full on Power and Maximum (dev.meta.ai)
  7. Mark Zuckerberg, X (@finkd), Zuckerberg's announcement with the benchmark table, (x.com)
  8. YouTube (AI at Meta), “Introducing Muse Wonder & Muse Wonder Full: frontier models for agentic tasks”, Official launch video, (youtube.com)
  9. Artificial Analysis, “Muse Wonder Full (max) - Intelligence, Performance & Price Analysis”, Index 56, 4th of 690; no public API (checked Oct. 14) (artificialanalysis.ai)
  10. Artificial Analysis, “Muse Wonder (max) - Intelligence, Performance & Price Analysis”, Muse Wonder scores 52 on the index (artificialanalysis.ai)
  11. Vals AI, Vals Index 64.57%, 5th of 48, (vals.ai)
  12. Vals AI, X (@ValsAI), "Meta now outranks OpenAI on the Vals Index" card, (x.com)
  13. Arena (LMArena), Text leaderboard: muse-wonder-max 2nd, 1,944 votes, (arena.ai)
  14. Cheeky Pint (YouTube), “Muse Wonder: Zuckerberg and Wang’s $100B AI”, Zuckerberg's check remark (00:53:11 and cold open); video on slide 7, (youtube.com)
  15. Cheeky Pint (Substack), “Muse Wonder: Zuckerberg and Wang’s $100B AI”, Official transcript, (cheekypint.substack.com)
  16. Meta (Investor Relations), “Meta Reports Second Quarter 2026 Results”, 2026 capital spending outlook of $130-145 billion, (investor.atmeta.com)
  17. CNBC, “Meta debuts Muse Wonder AI model in latest test of users’ willingness to pay”, Wang interview; subscription prices; Muse agent privacy complaints and Meta's response, (cnbc.com)
  18. VentureBeat, “Meta’s new Muse Wonder and Muse Wonder Full models achieve frontier performance, but only the weaker one is available via API”, API and subscription details; Power $50, Maximum $200, (venturebeat.com)
  19. TechCrunch, “Meta releases its first frontier model, Muse Wonder”, Wonder Full not on the API; Prometheus data center, (techcrunch.com)
  20. Decrypt, “Meta’s New Muse Wonder Finally Puts Zuckerberg in the Frontier AI Race”, Zuckerberg's Threads post; closed weights and privacy criticism, (decrypt.co)
  21. Futurum Group, “Meta Muse Wonder Reaches the Frontier. Can Meta Sell It?”, Analyst take; Gemini 4 Argon missing from the table, (futurumgroup.com)
  22. Wikimedia Commons, “File:Mark Zuckerberg (2025).jpg”, Cover photo: Jeff Sainlar / Meta, CC BY-SA 4.0, (commons.wikimedia.org)

About this story

Spotted an error in this story? [email protected] · Instagram

Short link: thejarrus.com/en/meta-muse-wonder

This story in Turkish: Meta, en yetenekli modelleri Muse Wonder ve Muse Wonder Full’u tanıttı

On the same topic