Meta 推出首款前沿模型 Muse Wonder，對標 GPT-6 Astra 與 Claude
Meta releases its first frontier model, Muse Wonder
摘要分析
2026 年 10 月 12 日 Meta 超級智慧實驗室發布 Muse Wonder 與 Muse Wonder Full，是 2025 年 8 月 Alexandr Wang 重組 Meta AI 後首款前沿等級模型。Meta 模型卡宣稱其基準成績可比 Anthropic Fable 5.1 與 OpenAI GPT-6 Astra。API 定價每百萬 token 輸入 4 美元、輸出 20 美元，與 GPT-6 Sol、Claude Opus 5.5 同級；Wonder Full 僅限 Muse Code 訂閱戶。兩款模型驅動 Muse 代理，主打保留並整合使用者資訊的「主動記憶」，自稱 MemoryBench 領先，訓練於俄亥俄州 Prometheus 資料中心。
- Meta 首款前沿模型，宣稱對標 Fable 5.1 與 GPT-6 Astra
- API 定價 4／20 美元，與主要對手同級
- Wonder Full 僅限 Muse Code 訂閱戶
- 主打「主動記憶」能力，MemoryBench 自稱領先
- 訓練於 2026 年啟用的 Prometheus 資料中心
On October 12, 2026, Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full, the company's first attempt at a frontier-class model since Alexandr Wang restructured Meta's AI efforts in August 2025. The Wonder models are larger than the lightweight Muse Spark and, per Meta's model card, post benchmark scores comparable to Anthropic's Fable 5.1 and OpenAI's GPT-6 Astra. Through the Meta Model API, Wonder costs $4 per million input tokens and $20 per million output tokens, roughly in line with GPT-6 Sol and Claude Opus 5.5; Wonder Full is reserved for Muse Code subscribers and not served via API. Both models power the Muse agent and emphasize 'active memory', the ability to retain and synthesize user-provided information, where Meta claims the lead on the MemoryBench benchmark. The models were trained at Meta's Prometheus data center in Ohio, which came online earlier in 2026.
- Meta's first frontier-class model, benchmarked against Fable 5.1 and GPT-6 Astra
- API pricing $4/$20 per million tokens, in line with rivals
- Wonder Full limited to Muse Code subscribers
- Emphasizes 'active memory'; claims MemoryBench lead
- Trained at the Prometheus data center that came online in 2026
關鍵字：MetaMuse Wonder前沿模型AI 記憶模型競爭
來源原文：https://techcrunch.com/2026/10/12/meta-releases-its-first-frontier-model-muse-wonder/
本頁為報導內容的摘要與分析，版權屬原媒體所有；請以原文為準。