[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-opus-5-fewer-refusals-ship-faster-zh":3,"article-related-opus-5-fewer-refusals-ship-faster-zh":30,"series-model-release-4fce0081-1ca3-44a1-91a7-1756615c769e":75},{"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":22,"views":26,"created_at":27,"published_at":28,"topic_cluster_id":29},"4fce0081-1ca3-44a1-91a7-1756615c769e","opus-5-fewer-refusals-ship-faster-zh","Opus 5 讓你少碰拒答","\u003Cp data-speakable=\"summary\">以前模型一拒答就斷線，現在 Opus 5 能把\u003Ca href=\"\u002Fnews\u002Fopenai-incident-postmortem-security-template-zh\">安全\u003C\u002Fa>檢查變成可用輸出。\u003C\u002Fp>\u003Cp>我最近一直在看 \u003Ca href=\"\u002Ftag\u002Fanthropic\">Anthropic\u003C\u002Fa> 的模型更新，老實講，前幾代我用得有點煩。不是它不強，是它太容易在我最需要它出手的地方踩煞車：安全分類器一抖，回覆就變成拒答；我想做的是產品流程，它回我的是政策邊界。你如果也做過 \u003Ca href=\"\u002Ftag\u002Fagent\">agent\u003C\u002Fa>、內部工具、客服輔助或 \u003Ca href=\"\u002Ftag\u002Fcode-review\">code review\u003C\u002Fa>，大概懂那種火氣：模型明明看得懂，卻硬是把你丟回起點。\u003C\u002Fp>\u003Cp>這次我被拉回來看，是因為 \u003Ca href=\"https:\u002F\u002Ftechcrunch.com\u002F2026\u002F07\u002F24\u002Fanthropic-launches-opus-5\u002F\">TechCrunch 的 Opus 5 報導\u003C\u002Fa>，再加上 Anthropic 自己把 \u003Ca href=\"https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fbuild-with-claude\u002Fsafety\u002Fautomatic-fallbacks\">Automatic Fallbacks\u003C\u002Fa> 放進文件。我不是被行銷詞打動，我是看到一個很實際的訊號：他們終於在處理「拒答後怎麼不要讓使用者整個卡死」這件事。這比模型名字聽起來多厲害，重要太多。\u003C\u002Fp>\u003Ch2>我在意的不是更會講，是更少卡住\u003C\u002Fh2>\u003Cblockquote>“Opus 5 was much stronger at verifying its work and iterating carefully until it succeeds.”\u003C\u002Fblockquote>\u003Cp>這句話我覺得很誠實。它沒有在吹什麼花俏能力，直接講模型會不會把事情做完。翻譯一下就是：它不是只會給你一個看起來漂亮的答案，而是會自己回頭檢查、補洞、重試，直到能交差。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785045786200-pciu.png\" alt=\"Opus 5 讓你少碰拒答\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>我自己最怕的就是那種「\u003Ca href=\"\u002Fnews\u002Fopenai-anthropic-dual-dominance-google-falls-behind-zh\">第一\u003C\u002Fa>句很像對的」模型。它一開始講得很順，結果中間漏了一個條件，後面就一路錯到底。你如果拿它做 agent，錯一次常常不是錯一次，是連鎖錯。這種時候，會驗證自己、會重試、會慢一點但把事做完的模型，反而比較像能上線的東西。\u003C\u002Fp>\u003Cp>我之前做過一個內部資料整理工具，模型很會寫摘要，但每次遇到多步驟抽取就開始亂補。使用者看起來有收到東西，實際上資料欄位一半是猜的。後來我才懂，能力不是只有「答得像」，而是「答得能收尾」。\u003C\u002Fp>\u003Cp>實操上我會這樣用：\u003C\u002Fp>\u003Cul>\u003Cli>把 Opus 5 放在需要多輪修正的任務：規劃、debug、資料轉換、工具調用。\u003C\u002Fli>\u003Cli>要求它在最後輸出前列出檢查步驟，尤其是 code 或結構化資料。\u003C\u002Fli>\u003Cli>評估時別只看一次答對率，還要看它能不能在第二輪自己修回來。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>便宜一點，才真的能拿來當預設\u003C\u002Fh2>\u003Cp>Anthropic 這次把 Opus 5 描述成更便宜、限制更少的選擇。這點我很買單，因為大模型真正麻煩的地方常常不是 \u003Ca href=\"\u002Ftag\u002Ftoken\">token\u003C\u002Fa> 費，而是你得替它補一堆流程洞。你本來以為自己在買算力，結果最後在養路由器、重試器、例外處理和客服工單。\u003C\u002Fp>\u003Cp>如果一個更強的模型，還同時比較便宜、比較不容易拒答，那它就不該只是「高級選項」，而該變成預設。這會直接改掉你的架構設計：你不用為了省錢硬切到小模型，也不用因為怕拒答就先塞一堆保守提示詞把模型綁死。\u003C\u002Fp>\u003Cp>我以前很常看到團隊把模型分層分到很累：A 類走便宜模型，B 類走大模型，C 類再加一層人工審核。結果真正的成本不是推理費，是維護這套分流規則的人力。模型一旦比較穩，很多這些補丁就能少一點。\u003C\u002Fp>\u003Cp>實操上我會這樣調整：\u003C\u002Fp>\u003Cul>\u003Cli>先檢查你現在的 routing 規則是不是過度保守，動不動就降級。\u003C\u002Fli>\u003Cli>看整體成功率，不要只看單次成本。\u003C\u002Fli>\u003Cli>把「因拒答而重試」當成一個要量化的指標，別讓它躲在客服抱怨裡。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>安全檢查還在，但不要每次都把路堵死\u003C\u002Fh2>\u003Cp>Anthropic 沒有假裝安全層不見了。文件和報導都提到，像 exploit 生成、滲透測試這類偏攻擊性的內容，還是會被擋；但他們也說，Opus 5 觸發這些分類器的頻率會比前一代低很多。數字我直接照文件脈絡講，重點是方向：少一點誤判，多一點可用性。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785045785089-kyra.png\" alt=\"Opus 5 讓你少碰拒答\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這件事我覺得很實際。安全層的存在我不反對，反而是必需品；我反對的是那種一刀切，明明是防禦性分析，卻被當成攻擊指令。你如果在做 source code review、\u003Ca href=\"\u002Fnews\u002Fai-regulation-india-business-risk-2026-zh\">風險\u003C\u002Fa>掃描、內部紅隊輔助，模型應該先理解意圖，再決定要不要踩煞車。\u003C\u002Fp>\u003Cp>我之前做過一個資安審查工具，團隊明明是在查防禦面問題，模型卻一直把 prompt 當成可疑攻擊流程。最後使用者學會了怎麼繞詞，結果 prompt 變得更長、更假、更難維護。那不是安全，那是把人逼去寫暗號。\u003C\u002Fp>\u003Cp>實操上我會這樣設計：\u003C\u002Fp>\u003Cul>\u003Cli>把防禦性工作和攻擊性工作分開寫 prompt，不要混在一起。\u003C\u002Fli>\u003Cli>一開始就講清楚用途：是分析、審查、偵錯，不是產生攻擊步驟。\u003C\u002Fli>\u003Cli>把 classifier 觸發次數記錄下來，觀察到底是安全真的生效，還是在亂擋。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>Automatic Fallbacks 才是能上線的那段\u003C\u002Fh2>\u003Cp>我覺得這次最值得抄的，不是模型本身，而是 \u003Ca href=\"https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fbuild-with-claude\u002Fsafety\u002Fautomatic-fallbacks\">Automatic Fallbacks\u003C\u002Fa>。意思很簡單：當安全分類器擋下來時，不一定直接丟錯誤，而是可以轉去比較弱、比較保守的模型，先給使用者一個還能用的結果。\u003C\u002Fp>\u003Cp>翻譯一下就是：別讓一次拒答變成整個流程死亡。這個想法很土，但很對。因為大多數產品根本不需要模型在每個瞬間都完美通過政策檢查，它們需要的是使用者不要卡住。就算答案比較保守、比較短、比較不像大模型，也比空白頁強。\u003C\u002Fp>\u003Cp>我很喜歡這種做法，因為它承認一件事：拒答有時候是對的，但把拒答包成硬錯誤，通常只是 UX 很爛。你如果能把它降級成可控 fallback，產品就比較像產品，不像 demo。\u003C\u002Fp>\u003Cp>我以前做過一個客服輔助工具，最痛的不是模型答錯，而是它直接不回。客服人員面對空白結果，只能重送、改寫、猜 prompt。後來我們加了 fallback，哪怕答案比較保守，至少能先把工單往前推。那種「先有東西，再慢慢修」的節奏，真的差很多。\u003C\u002Fp>\u003Cp>實操上我會這樣做：\u003C\u002Fp>\u003Cul>\u003Cli>對使用者介面來說，fallback 預設比硬拒答更好。\u003C\u002Fli>\u003Cli>fallback 模型只做低風險任務，不要讓它偷偷升級成另一個大腦。\u003C\u002Fli>\u003Cli>把 fallback 事件寫進 log，否則你根本不知道主模型是不是太常踩線。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>我會先看三個數字，再決定要不要全切\u003C\u002Fh2>\u003Cp>我不會看到新模型就整套搬過去。這種事最容易翻車。真正該看的只有三個：拒答率、fallback 品質、以及它到底有沒有真的減少後續修正。這三個指標如果沒變好，再強的敘事都只是換包裝。\u003C\u002Fp>\u003Cp>也就是說，評估不要做成儀式，要做成營運。你要看的是：第一次或第二次回合，能不能拿到可用答案；fallback 有沒有保住原本意圖；安全層有沒有保護系統，同時又沒把正常工作都擋掉。\u003C\u002Fp>\u003Cp>我自己會把評估拆成兩層。第一層看產品：使用者有沒有卡住、客服有沒有爆、任務有沒有完成。第二層看工程：路由邏輯有沒有太複雜、例外處理有沒有越寫越厚、日誌有沒有能回頭追。這樣你才知道問題是模型、策略，還是你自己架太爛。\u003C\u002Fp>\u003Cp>如果這三個數字都往好的方向走，Opus 5 才值得當預設。否則它只是另一個名字比較好記的選項。\u003C\u002Fp>\u003Ch2>可抄的模板\u003C\u002Fh2>\u003Cpre>\u003Ccode># Anthropic Opus 5 production playbook\n\n## 1) Primary model\nUse Opus 5 first for:\n- multi-step reasoning\n- code generation and repair\n- defensive security analysis\n- tool-heavy agent workflows\n- tasks that need self-verification\n\n## 2) Prompt style\nState intent early and plainly:\n- \"This is for defensive analysis only.\"\n- \"Review the code for vulnerabilities.\"\n- \"Do not generate exploit instructions.\"\n- \"Check your work before answering.\"\n\n## 3) Routing logic\n1. Send the request to Opus 5.\n2. If the safety layer blocks it, do not show a raw error first.\n3. Route to a smaller, safer fallback model.\n4. Return the fallback result if it still fits the user’s intent.\n5. If the request is clearly malicious, return a policy-safe refusal.\n\n## 4) Fallback policy\nFallback is allowed only when:\n- the request is safe but over-triggered\n- a partial answer is better than no answer\n- the fallback can preserve the original intent\n\nFallback is not allowed when:\n- the request is offensive security guidance\n- the request is obviously malicious\n- the fallback would hide a real policy violation\n\n## 5) Logging\nLog these fields:\n- request id\n- original prompt category\n- safety trigger reason\n- primary model result\n- fallback model used\n- usable-answer flag\n- manual-correction flag\n\n## 6) Evaluation checklist\nTrack weekly:\n- refusal rate\n- fallback rate\n- usable-answer rate\n- manual correction rate\n- time to resolution\n- user-reported dead ends\n\n## 7) Minimal implementation sketch\nrequest -> Opus 5\n  if allowed:\n    return response\n  if safety-triggered and fallback enabled:\n    route to smaller model\n    return fallback response\n  if still unsafe:\n    return policy-safe refusal with a short explanation\n\n## 8) Team rule\nTreat safety failures as routing events.\nTreat hard refusals as the last resort, not the default UX.\u003C\u002Fcode>\u003C\u002Fpre>\u003Cp>我拆這篇的原始來源是 \u003Ca href=\"https:\u002F\u002Ftechcrunch.com\u002F2026\u002F07\u002F24\u002Fanthropic-launches-opus-5\u002F\">TechCrunch\u003C\u002Fa>，以及 Anthropic 的 \u003Ca href=\"https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fbuild-with-claude\u002Fsafety\u002Fautomatic-fallbacks\">Automatic Fallbacks\u003C\u002Fa> 文件。上面關於怎麼落地、怎麼評估、怎麼寫模板，是我自己把這些材料翻成開發者能直接拿去用的版本。\u003C\u002Fp>","我拆 Anthropic Opus 5 的實用打法，重點是怎麼用 fallback 把安全拒答變成可用輸出。","techcrunch.com","https:\u002F\u002Ftechcrunch.com\u002F2026\u002F07\u002F24\u002Fanthropic-launches-opus-5\u002F",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785045786200-pciu.png","model-release","zh","43cd3860-e32d-4585-94c3-c9de55a8dad9",[17,18,19,20,21],"Anthropic","Opus 5","fallback","safety classifier","agent workflow",[23,24,25],"Opus 5 的價值在於更少拒答、更多可用輸出。","Automatic Fallbacks 把安全拒答改成可控降級。","上線前先量拒答率、fallback 品質和可用答案率。",0,"2026-07-26T06:02:43.075148+00:00","2026-07-26T06:02:43.061+00:00","d17e8d03-aed6-4aa1-add8-835c042bb5a9",{"tags":31,"relatedLang":34,"relatedPosts":38},[32],{"name":17,"slug":33},"anthropic",{"id":15,"slug":35,"title":36,"language":37},"opus-5-fewer-refusals-ship-faster-en","Opus 5 lets you ship with fewer refusals","en",[39,45,51,57,63,69],{"id":40,"slug":41,"title":42,"cover_image":43,"image_url":43,"created_at":44,"category":13},"74cd6162-e59b-45f1-9fb0-d315b14b1da7","claude-opus-5-undercuts-fable-5-price-zh","Claude Opus 5 以更低價格搶企業單","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784982778740-4mwu.png","2026-07-25T12:32:32.937725+00:00",{"id":46,"slug":47,"title":48,"cover_image":49,"image_url":49,"created_at":50,"category":13},"241f7378-04d7-4fdd-903d-5c32afbea477","openai-gpt-5-6-pricing-tiers-zh","OpenAI 列出 GPT-5.6 三檔定價","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784811764336-hr3z.png","2026-07-23T13:02:23.344674+00:00",{"id":52,"slug":53,"title":54,"cover_image":55,"image_url":55,"created_at":56,"category":13},"20c23f9d-2ddf-43b1-97ba-f97a494b006e","gemini-3-6-flash-efficiency-over-hype-zh","Gemini 3.6 Flash 證明 Google 把效率放在 hype 前面","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784725369011-xwnw.png","2026-07-22T13:02:20.322757+00:00",{"id":58,"slug":59,"title":60,"cover_image":61,"image_url":61,"created_at":62,"category":13},"0de8d105-5bda-4eab-8cba-fb6f9515ec78","kimi-k3-grok-build-code-analysis-zh","Kimi K3讀懂82萬行 Grok Build 代碼","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784711003170-xiom.png","2026-07-22T09:02:38.007579+00:00",{"id":64,"slug":65,"title":66,"cover_image":67,"image_url":67,"created_at":68,"category":13},"98e69dcf-3061-407f-b821-e37370180463","gpt-5-6-three-variants-lower-token-costs-zh","GPT-5.6 三版本登場，Token 成本更低","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784282584596-z26c.png","2026-07-17T10:02:36.272941+00:00",{"id":70,"slug":71,"title":72,"cover_image":73,"image_url":73,"created_at":74,"category":13},"0b89c453-80d5-4b7e-b183-d274c1907a0b","gpt-56-sol-terra-luna-digitalocean-ai-zh","GPT-5.6 三模型上線 DigitalOcean","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784273583159-7xfk.png","2026-07-17T07:32:36.367538+00:00",[76,81,86,91,96,101,106,111,116,121],{"id":77,"slug":78,"title":79,"created_at":80},"58b64033-7eb6-49b9-9aab-01cf8ae1b2f2","nvidia-rubin-six-chips-one-ai-supercomputer-zh","NVIDIA Rubin 把六顆晶片塞進 AI 機櫃","2026-03-26T07:18:45.861277+00:00",{"id":82,"slug":83,"title":84,"created_at":85},"0dcc2c61-c2a6-480d-adb8-dd225fc68914","march-2026-ai-model-news-what-mattered-zh","2026 年 3 月 AI 模型新聞重點","2026-03-26T07:32:08.386348+00:00",{"id":87,"slug":88,"title":89,"created_at":90},"214ab08b-5ce5-4b5c-8b72-47619d8675dd","why-small-models-are-winning-on-device-ai-zh","小模型為何吃下裝置端 AI","2026-03-26T07:36:30.488966+00:00",{"id":92,"slug":93,"title":94,"created_at":95},"785624b2-0355-4b82-adc3-de5e45eecd88","midjourney-v8-faster-images-higher-costs-zh","Midjourney V8 變快了，也變貴了","2026-03-26T07:52:03.562971+00:00",{"id":97,"slug":98,"title":99,"created_at":100},"9e1044b4-946d-47fe-9e2a-c2ee032e1164","xiaomi-mimo-v2-pro-1t-moe-agents-zh","小米 MiMo-V2-Pro 登場：1T MoE 模型","2026-03-28T03:06:19.002353+00:00",{"id":102,"slug":103,"title":104,"created_at":105},"c4b6186f-bd84-4598-997e-c6e31d543c0d","cursor-composer-2-agentic-coding-model-zh","Cursor Composer 2 走向代理式寫碼","2026-03-28T03:13:06.422716+00:00",{"id":107,"slug":108,"title":109,"created_at":110},"e112e76f-ec3b-408f-810e-e93ae21a888a","apple-siri-gemini-distilled-models-zh","Apple Siri 牽手 Gemini 的真相","2026-03-29T04:52:57.886544+00:00",{"id":112,"slug":113,"title":114,"created_at":115},"c679b51f-194a-463b-87fc-7695256ff752","mimo-v2-pro-vs-omni-vs-flash-2026-zh","MiMo V2 Pro、Omni、Flash 怎麼選","2026-04-02T01:18:43.576128+00:00",{"id":117,"slug":118,"title":119,"created_at":120},"3b988fd7-6749-4f01-ba25-c0ad7486dc31","z-ai-glm-5v-turbo-design2code-claude-zh","GLM-5V-Turbo 在 Design2Code 贏了…","2026-04-02T04:03:36.31741+00:00",{"id":122,"slug":123,"title":124,"created_at":125},"975a7aef-030e-41a6-9401-1c6a342be68e","april-2026-ai-model-releases-zh","2026年4月 AI 模型更新追蹤","2026-04-02T08:45:33.308563+00:00"]