Muse Wonder評測:Meta首款前沿模型,最強版鎖訂閱 | Muse Wonder Review: Meta's Frontier Model, Locked Tier
By Kit 小克 | AI Tool Observer | 2026-10-13
🇹🇼 Muse Wonder評測:Meta首款前沿模型,最強版鎖訂閱
Meta在10月12日發布了公司首款「前沿級」(frontier-class)模型 Muse Wonder 與升級版 Muse Wonder Full,祖克柏稱這是邁向「個人超級智能」的重要一步。但魔鬼藏在定價與跑分細節裡:真正強的那一版,你得先掏錢訂閱才能碰到。
兩個版本,兩種待遇:API開放 vs 訂閱限定
標準版 Muse Wonder 透過API開放給所有開發者,定價每百萬輸入token $4、輸出token $20,快取輸入降到$0.2,網頁搜尋功能每千次查詢加收$2.5,context window給到100萬token,速率上限每分鐘3000次請求、400萬token。這個價位帶跟Claude系列幾乎對齊,不算便宜但也在合理範圍。
問題在升級版。Muse Wonder Full 被Meta自己定義為「真正的前沿模型」,卻完全不開放API,只給Muse Code訂閱戶使用:Power層每月$50、Maximum層每月$200。換句話說,你沒辦法單純花token費用用到最強版本,得先綁訂閱才行。
跑分不差,但宣稱經不起重新核算
第三方測試機構Artificial Analysis的Intelligence Index給 Muse Wonder Full 56分,追平Claude Sonnet 5.5,並超越GPT-6 Astra、Gemini 4 Argon與Claude Fable 5.1;標準版Wonder則是52分。看起來很硬,但Claude Opus 5.5仍以58分穩居上風。
更尷尬的是,Meta宣稱 Muse Wonder Full 在多項「持續型代理任務」測試(包括DeepSWE)上打敗Opus 5.5,可是外部用Meta自己公布的原始數據重新核算,結果對不上。有分析進一步指出,在日常編碼任務裡,Wonder Full並沒有比免費可得的標準版多完成專案、少一次人工修正,或跑得更快——如果你已經在用Claude Code或Codex,額外訂閱Wonder Full可能換不到實際生產力提升。
小克的實測建議
這次發布的商業意義大於技術意義:Meta等於正式向市場宣告「我們也擠進frontier等級」,股價、話題都有了。但對一般開發者來說,真正划算的做法是先用API版Muse Wonder跑跑你自己的任務,看複雜度是否真的卡在標準版做不到的地方,再決定要不要多付$50~$200訂閱Full版本——目前公開的證據顯示,大部分日常用途不需要。
好不好用,試了才知道。
🇺🇸 Muse Wonder Review: Meta's Frontier Model, Locked Tier
Meta shipped its first "frontier-class" model on October 12: Muse Wonder and a beefed-up Muse Wonder Full, which Mark Zuckerberg called a step toward "personal superintelligence." But the catch is in the pricing and the benchmarks — the version that's actually competitive is locked behind a subscription.
Two Tiers, Two Different Deals: API vs. Subscription-Only
The standard Muse Wonder is open to all developers via API at $4 per million input tokens and $20 per million output tokens, with cached input discounted to $0.20/M and web search billed at $2.50 per 1,000 queries. It ships with a 1-million-token context window and rate limits of 3,000 requests and 4 million tokens per minute. That pricing lines up closely with Claude's tier — not cheap, but not unreasonable either.
The problem is the upgrade. Muse Wonder Full, which Meta itself calls the "real" frontier model, isn't available via API at all. It's reserved exclusively for Muse Code subscribers: $50/month for the Power tier, $200/month for Maximum. You can't just pay per token to touch the best version — you have to commit to a subscription first.
Solid Benchmarks, Shaky Claims Under Re-Checking
Independent benchmarking firm Artificial Analysis scored Muse Wonder Full at 56 on its Intelligence Index — tying Claude Sonnet 5.5 and beating GPT-6 Astra, Gemini 4 Argon, and Claude Fable 5.1. The standard Wonder scored 52. That's respectable, but Claude Opus 5.5 still leads at 58.
More awkward: Meta claimed Muse Wonder Full beats Opus 5.5 on several sustained-agent-work benchmarks, including DeepSWE — but when outside analysts recomputed the numbers using Meta's own published data, the claims didn't reproduce. Further testing found that for routine coding tasks, Wonder Full didn't finish more projects, need fewer manual fixes, or run faster than the freely available standard Wonder. If you're already running Claude Code or Codex, bolting on a Wonder Full subscription may not buy you any real productivity gain.
Kit's Honest Take
This launch matters more as a business signal than a technical one — Meta is telling the market it has arrived at frontier-tier performance, and that narrative alone moves headlines. For actual developers, the smart move is to try the API-accessible Muse Wonder on your own workload first, see whether you genuinely hit a wall the standard tier can't clear, and only then weigh the $50–$200/month for Full. Based on the evidence so far, most everyday use cases don't need it.
You won't know until you try it.
Sources / 資料來源
- VentureBeat: Muse Wonder Full 只有弱版開放API
- FelloAI: Muse Wonder 跑分、價格與訂閱陷阱
- OrcaRouter: Muse Wonder Full 持續型代理任務實測
延伸閱讀 / Related Articles
- QGIS評測:開源地圖軟體封殺AI程式碼貢獻 | QGIS Review: Open-Source GIS Bans AI Code Contributions
- Copilot多模型評測:微軟砸20億美元導入Claude | Copilot Multi-Model Review: Microsoft Pays Anthropic $2B
- Claude假線報評測:AI亂報兇殺案,白宮要求究責 | Claude False Tip Review: AI Agent's Fake Police Report
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言