前沿模型智能接近,单位任务成本相差 3.8 倍
五款前沿模型的智能水平相差不到 5 分,但单位任务成本相差 3.8 倍。Google 的 Argon 以每任务 1.99 美元处于最低位,比最贵的 Fable 5.1 低 74%,比 Opus 5.5 低 67%、比 Muse Wonder 低 56%、比 Astra 低 39%。作者认为这与 Google 自研 TPU 十余年的积累有关,并援引 Microsoft Maia 300 与 OpenAI 推理芯片 Jalapeño 的公开数据,判断随着每瓦智能每六个月翻倍,token 价格会继续下行,前沿模型利润率仍可维持在 40% 至 70%。
y-leading speed and efficiency in AI inference”, August 25, 2026, reporting results measured on SemiAnalysis’s InferenceX suite with OpenAI engineers present, across GPT-OSS 120B, DeepSeek R1 670B & Kimi K2.5 1T. OpenAI normalized to published package power & reports Jalapeño’s sustained draw at or below 550W against a 700W rating. Three caveats apply : the comparison is against Blackwell-generation GB200/GB300 rather than the newer Vera Rubin platform, which also uses HBM4 & began shipping to c
匹配字段正文

