[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-llm-routing-benchmark-38-tasks-15-models-zh":3,"article-related-llm-routing-benchmark-38-tasks-15-models-zh":35,"series-industry-0c3a0e19-e3c7-48ae-8f7b-198ca0911957":78},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":27,"views":31,"created_at":32,"published_at":33,"topic_cluster_id":34},"0c3a0e19-e3c7-48ae-8f7b-198ca0911957","llm-routing-benchmark-38-tasks-15-models-zh","5 款 LLM 的實戰路由結論","\u003Cp>\u003Ca href=\"\u002Fnews\u002Fopen-source-agent-stacks-seven-layers-2026-zh\">2026\u003C\u002Fa> 年做實際工作時，\u003Ca href=\"\u002Ftag\u002Fllm\">LLM\u003C\u002Fa> 到底該選哪一個？\u003C\u002Fp>\u003Cp data-speakable=\"summary\">這份基準測試用 15 個\u003Ca href=\"\u002Fnews\u002Fkimi-k3-model-hype-into-harness-work-zh\">模型\u003C\u002Fa>、38 項真實任務，整理出主力、低成本、結構化輸出和本地免費模型的分工。\u003C\u002Fp>\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>項目\u003C\u002Fth>\u003Cth>品質\u003C\u002Fth>\u003Cth>通過率\u003C\u002Fth>\u003Cth>中位時間\u003C\u002Fth>\u003Cth>總成本\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fclaude\">Claude Sonnet 4.6\u003C\u002Fa>\u003C\u002Ftd>\u003Ctd>100%\u003C\u002Ftd>\u003Ctd>38\u002F38\u003C\u002Ftd>\u003Ctd>4.6s\u003C\u002Ftd>\u003Ctd>$0.20\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fclaude\">Claude Opus 4.6\u003C\u002Fa>\u003C\u002Ftd>\u003Ctd>100%\u003C\u002Ftd>\u003Ctd>38\u002F38\u003C\u002Ftd>\u003Ctd>4.1s\u003C\u002Ftd>\u003Ctd>$0.69\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fwww.minimax.io\u002F\">MiniMax M2.5\u003C\u002Fa>\u003C\u002Ftd>\u003Ctd>98.6%\u003C\u002Ftd>\u003Ctd>38\u002F38\u003C\u002Ftd>\u003Ctd>15.9s\u003C\u002Ftd>\u003Ctd>$0.07\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fai.google.dev\u002Fgemini-api\">Gemini 2.5 Flash\u003C\u002Fa>\u003C\u002Ftd>\u003Ctd>97.1%\u003C\u002Ftd>\u003Ctd>35\u002F38\u003C\u002Ftd>\u003Ctd>1.1s\u003C\u002Ftd>\u003Ctd>$0.003\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fopenai.com\u002Fresearch\u002Fopenai-oss\">GPT-oss-20b\u003C\u002Fa>\u003C\u002Ftd>\u003Ctd>98.3%\u003C\u002Ftd>\u003Ctd>37\u002F38\u003C\u002Ftd>\u003Ctd>4.1s\u003C\u002Ftd>\u003Ctd>$0.00\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003Ch2>1. Sonnet 是最穩的主力\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fclaude\">Claude Sonnet 4.6\u003C\u002Fa> 是這批模型裡最平衡的選擇，38\u002F38 全過，總成本只有 $0.20，還能把中位回應時間壓在 4.6 秒。對多數團隊來說，這代表它可以直接當預設路由。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784422962336-b8ht.png\" alt=\"5 款 LLM 的實戰路由結論\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>它的價值不只是準，而是準得夠便宜。若你的工作包含程式、抽取、摘要與一般寫作，Sonnet 很適合放在「大多數任務先送它」的位置。\u003C\u002Fp>\u003Cul>\u003Cli>38\u002F38 任務通過\u003C\u002Fli>\u003Cli>172.5\u002F172.5 分\u003C\u002Fli>\u003Cli>4.6 秒中位回應\u003C\u002Fli>\u003Cli>$0.20 總成本\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>2. Opus 不是更划算的升級\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fclaude\">Claude Opus 4.6\u003C\u002Fa> 同樣拿到 38\u002F38，但總成本升到 $0.69，沒有在這組任務裡展現出足以說服人的品質差距。也就是說，它贏在名義定位，不一定贏在實際回報。\u003C\u002Fp>\u003Cp>這不代表 Opus 沒用，而是它更像高難度任務的保險。當你已經知道問題很棘手，或需要更保守的推理空間時，它才\u003Ca href=\"\u002Fnews\u002Fopen-source-ai-agent-frameworks-compared-langfuse-zh\">比較\u003C\u002Fa>值得加價。\u003C\u002Fp>\u003Cul>\u003Cli>38\u002F38 任務通過\u003C\u002Fli>\u003Cli>4.1 秒中位回應\u003C\u002Fli>\u003Cli>與 Sonnet 同分\u003C\u002Fli>\u003Cli>成本最高\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>3. MiniMax 最適合嚴格格式輸出\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fwww.minimax.io\u002F\">MiniMax M2.5\u003C\u002Fa> 的亮點不是只看分數，而是它很少亂加包裝文字。對要接 \u003Ca href=\"\u002Ftag\u002Fapi\">API\u003C\u002Fa>、要進 parser、要直接寫入資料庫的流程來說，這種輸出紀律很重要。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784422962994-5zvn.png\" alt=\"5 款 LLM 的實戰路由結論\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>它在這份測試裡拿到 98.6% 品質與 38\u002F38 通過率，成本也只有 $0.07。若你的重點是結構化資料而不是長篇推理，這是值得先測的模型。\u003C\u002Fp>\u003Cul>\u003Cli>98.6% 品質\u003C\u002Fli>\u003Cli>38\u002F38 通過\u003C\u002Fli>\u003Cli>$0.07 總成本\u003C\u002Fli>\u003Cli>多數輸出為純 JSON\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>4. Flash 最省錢也最快\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fai.google.dev\u002Fgemini-api\">Gemini 2.5 Flash\u003C\u002Fa> 把速度與價格壓到很低：中位時間 1.1 秒，總成本只有 $0.003。它的 97.1% 品質不算最高，但對大量批次處理已經很夠用。\u003C\u002Fp>\u003Cp>如果任務是清洗、轉換、分類或其他規則明確的工作，Flash 很適合當高吞吐路由。它不是最強，但常常是最划算的那個。\u003C\u002Fp>\u003Cul>\u003Cli>97.1% 品質\u003C\u002Fli>\u003Cli>35\u002F38 通過\u003C\u002Fli>\u003Cli>1.1 秒中位回應\u003C\u002Fli>\u003Cli>$0.003 總成本\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>5. GPT-oss-20b 提供免費本地路線\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fopenai.com\u002Fresearch\u002Fopenai-oss\">GPT-oss-20b\u003C\u002Fa> 的驚喜在於它是本地跑、總成本 $0.00，卻還能拿到 98.3% 品質與 37\u002F38 通過率。對重視內網部署或固定成本控制的團隊，這很有吸引力。\u003C\u002Fp>\u003Cp>它的定位不是全能王，而是可用的免費工作馬。若你的任務不屬於深度推理型，本地模型已經可以進入正式流程，而不只是測試環境。\u003C\u002Fp>\u003Cul>\u003Cli>98.3% 品質\u003C\u002Fli>\u003Cli>37\u002F38 通過\u003C\u002Fli>\u003Cli>$0.00 總成本\u003C\u002Fli>\u003Cli>可本地部署\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>怎麼挑\u003C\u002Fh2>\u003Cp>想要一個主力模型，先選 Sonnet。想要最穩的結構化輸出，先測 MiniMax。想要最低延遲和最低成本，Flash 最有說服力。想要零邊際成本與本地控制，GPT-oss-20b 值得優先試。\u003C\u002Fp>\u003Cp>這份測試真正傳達的不是單一冠軍，而是路由思維。把簡單任務交給便宜模型，把難題留給高階模型，通常比押寶單一模型更有效。\u003C\u002Fp>","15 個模型、38 項真實任務顯示：Sonnet 適合主力，MiniMax 擅長結構化輸出，Flash 最省最快，GPT-oss-20b 可本地免費跑。","ianlpaterson.com","https:\u002F\u002Fianlpaterson.com\u002Fblog\u002Fllm-benchmark-2026-38-actual-tasks-15-models-for-2-29\u002F",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784422962336-b8ht.png","industry","zh","0eb521a5-8e5c-4825-884f-18f5f045bca7",[17,18,19,20,21,22,23,24,25,26],"LLM benchmark","routing","Claude Sonnet 4.6","Claude Opus 4.6","MiniMax M2.5","Gemini 2.5 Flash","GPT-oss-20b","real-world tasks","structured output","local model",[28,29,30],"Sonnet 是最均衡的主力選擇，適合多數通用工作。","MiniMax 對 JSON 與結構化輸出特別友好。","Flash 最便宜最快，GPT-oss-20b 則適合本地免費部署。",0,"2026-07-19T01:02:19.036094+00:00","2026-07-19T01:02:19.029+00:00","7aa69b8b-ff49-4d68-9e8b-f08e577b1239",{"tags":36,"relatedLang":37,"relatedPosts":41},[],{"id":15,"slug":38,"title":39,"language":40},"llm-benchmark-2026-38-real-tasks-15-models-en","15 LLMs on 38 tasks show routing beats one-model bets","en",[42,48,54,60,66,72],{"id":43,"slug":44,"title":45,"cover_image":46,"image_url":46,"created_at":47,"category":13},"ca7a1984-de05-4f98-b029-a96105ee6f9c","wuping-ai-yanjing-xianying-dai-ping-ar-en-zh","无屏AI眼镜先赢，带屏AR还得等芯片成熟","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784491387359-hw7j.png","2026-07-19T20:02:40.585372+00:00",{"id":49,"slug":50,"title":51,"cover_image":52,"image_url":52,"created_at":53,"category":13},"674c3974-912e-4fe3-a727-5d65ed85e2c4","nvidia-open-docs-defensive-concession-zh","Nvidia 的開源文檔不是開放，而是防禦性讓步","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784489575324-aolb.png","2026-07-19T19:32:30.65371+00:00",{"id":55,"slug":56,"title":57,"cover_image":58,"image_url":58,"created_at":59,"category":13},"35954623-589c-42aa-b871-979c98690225","wgu-anthropic-ai-native-credentialing-model-zh","WGU 與 Anthropic 押注 AI 原生證照","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784487769887-lbsn.png","2026-07-19T19:02:21.899746+00:00",{"id":61,"slug":62,"title":63,"cover_image":64,"image_url":64,"created_at":65,"category":13},"755fe1e6-ace4-4533-96aa-ee841f7666fc","agi-2026-five-facts-cut-through-noise-zh","2026 年 AGI 仍未定案的 5 個關鍵事實","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784440969793-3tr5.png","2026-07-19T06:02:21.135709+00:00",{"id":67,"slug":68,"title":69,"cover_image":70,"image_url":70,"created_at":71,"category":13},"fb7e9fae-1b07-4f17-9445-62c0ae5ae401","openai-staff-fund-rival-super-pac-zh","OpenAI 員工捐 21.5 萬美元挺反 AI PAC","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784419368821-l8ql.png","2026-07-19T00:02:21.179629+00:00",{"id":73,"slug":74,"title":75,"cover_image":76,"image_url":76,"created_at":77,"category":13},"2299155a-c8ca-40e1-9916-dd43f8f7a35f","open-source-agent-stacks-seven-layers-2026-zh","2026 開源 agent 堆疊拆成七層","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784403178240-b3c3.png","2026-07-18T19:32:32.446009+00:00",[79,84,89,94,99,104,109,114,119,124],{"id":80,"slug":81,"title":82,"created_at":83},"ee073da7-28b3-4752-a319-5a501459fb87","ai-in-2026-what-actually-matters-now-zh","2026 AI 真正重要的事","2026-03-26T07:09:12.008134+00:00",{"id":85,"slug":86,"title":87,"created_at":88},"83bd1795-8548-44c9-9a7e-de50a0923f71","trump-ai-framework-power-speech-state-preemption-zh","川普 AI 框架瞄準電力、言論與州權","2026-03-26T07:12:18.695466+00:00",{"id":90,"slug":91,"title":92,"created_at":93},"ea6be18b-c903-4e54-97b7-5f7447a612e0","nvidia-gtc-2026-big-ai-announcements-zh","NVIDIA GTC 2026 重點拆解","2026-03-26T07:14:26.62638+00:00",{"id":95,"slug":96,"title":97,"created_at":98},"4bcec76f-4c36-4daa-909f-54cd702f7c93","claude-users-spreading-out-and-getting-better-zh","Claude 用戶更分散，也更會用","2026-03-26T07:22:52.325888+00:00",{"id":100,"slug":101,"title":102,"created_at":103},"bd903b15-2473-4178-9789-b7557816e535","openclaw-raises-hard-question-for-ai-models-zh","OpenClaw 逼問 AI 模型價值","2026-03-26T07:24:54.707486+00:00",{"id":105,"slug":106,"title":107,"created_at":108},"eeac6b9e-ad9d-4831-8eec-8bba3f9bca6a","gap-google-gemini-checkout-fashion-search-zh","Gap 把結帳搬進 Gemini","2026-03-26T07:28:23.937768+00:00",{"id":110,"slug":111,"title":112,"created_at":113},"0740e53f-605d-4d57-8601-c10beb126f3c","google-pushes-gemini-transition-to-march-2026-zh","Google 把 Gemini 轉換延到 2026 年 3…","2026-03-26T07:30:12.825269+00:00",{"id":115,"slug":116,"title":117,"created_at":118},"e660d801-2421-4529-8fa9-86b82b066990","metas-llama-4-benchmark-scandal-gets-worse-zh","Meta Llama 4 分數風波又擴大","2026-03-26T07:34:21.156421+00:00",{"id":120,"slug":121,"title":122,"created_at":123},"183f9e7c-e143-40bb-a6d5-67ba84a3a8bc","accenture-mistral-ai-sovereign-enterprise-deal-zh","Accenture 攜手 Mistral AI 賣主權 AI","2026-03-26T07:38:14.818906+00:00",{"id":125,"slug":126,"title":127,"created_at":128},"191d9b1b-768a-478c-978c-dd7431a38149","mistral-ai-faces-its-hardest-year-yet-zh","Mistral AI 迎來最硬的一年","2026-03-26T07:40:23.716374+00:00"]