[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-claude-opus-45-gpt-parameters-estimate-zh":3,"article-related-claude-opus-45-gpt-parameters-estimate-zh":30,"series-research-838cb5fd-5651-49fb-9b4c-c2dbde25ca02":86},{"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":11,"views":27,"created_at":28,"published_at":29,"topic_cluster_id":11},"838cb5fd-5651-49fb-9b4c-c2dbde25ca02","claude-opus-45-gpt-parameters-estimate-zh","Claude Opus 4.5 和 GPT 到底多大","\u003Cp>大家常把前沿 \u003Ca href=\"\u002Fnews\u002Fai-maps-navigation-mcp-baidu-autonavi-tencent-zh\">AI\u003C\u002Fa> 想成「越大越猛」。但這件事，現在沒那麼直線了。\u003Ca href=\"https:\u002F\u002Fopenai.com\u002Findex\u002Fgpt-4\u002F\" target=\"_blank\" rel=\"noopener\">GPT-4\u003C\u002Fa> 曾被外界估到 1.6 兆參數。\u003Ca href=\"https:\u002F\u002Fopenai.com\u002Findex\u002Fgpt-4o-and-more-tools-to-chatgpt-free\u002F\" target=\"_blank\" rel=\"noopener\">GPT-4o\u003C\u002Fa> 的估計卻掉到 200B 到 300B。差很多，真的差很多。\u003C\u002Fp>\u003Cp>這也讓 \u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002F\" target=\"_blank\" rel=\"noopener\">Anthropic\u003C\u002Fa> 的 \u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Fclaude-4\" target=\"_blank\" rel=\"noopener\">Claude Opus\u003C\u002Fa> 變得更有意思。官方沒講參數數字，但市場早就不只看大小。大家更在意成本、延遲、吞吐量，還有模型到底能不能把活做好。\u003C\u002Fp>\u003Cp>講白了，參數數量不是全部。可它還是很有參考價值。因為它會直接影響訓練成本、伺服器成本，還有產品能不能大規模上線。對開發者來說，這比「聽起來很大」重要太多。\u003C\u002Fp>\u003Ch2>參數大小，為什麼還是很重要\u003C\u002Fh2>\u003Cp>參數數量是一個粗糙指標。可是它很常反映現實世界的帳單。模型越大，訓練通常越貴。推論時也更吃 GPU，尤其是 dense model。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775207388141-adee.png\" alt=\"Claude Opus 4.5 和 GPT 到底多大\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>如果一個模型接近 1.6T 參數，那它的運行壓力就很大。反過來，200B 到 300B 的模型，通常更容易壓低服務成本。這也是為什麼很多公司開始追求更精簡的架構。\u003C\u002Fp>\u003Cp>但你也不能只看數字。資料品質、訓練配方、MoE 路由、後訓練、工具調用，這些都會改變結果。說真的，有時候一個比較小的模型，實際用起來反而更順。\u003C\u002Fp>\u003Cul>\u003Cli>\u003Cstrong>GPT-4：\u003C\u002Fstrong>外界常估約 1.6T 參數\u003C\u002Fli>\u003Cli>\u003Cstrong>GPT-4o：\u003C\u002Fstrong>常見估計落在 200B 到 300B\u003C\u002Fli>\u003Cli>\u003Cstrong>Claude Opus：\u003C\u002Fstrong>官方沒公開參數數字\u003C\u002Fli>\u003Cli>\u003Cstrong>推論成本：\u003C\u002Fstrong>通常跟模型大小高度相關\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>公開線索到底透露了什麼\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fopenai.com\u002F\" target=\"_blank\" rel=\"noopener\">OpenAI\u003C\u002Fa> 沒公開 GPT-4 和 GPT-4o 的參數數字，所以外界只能從間接線索拼圖。研究者會看模型表現、基礎設施痕跡、以及和合作夥伴相關的公開資訊。這不是精準答案，但方向很清楚。\u003C\u002Fp>\u003Cp>GPT-4 長期被當成超大模型。1.6T 這個數字在技術圈流傳很久。到了 GPT-4o，敘事就變了。它更快，也更像是「夠大，但沒大到離譜」的那一類。\u003C\u002Fp>\u003Cp>這裡可以直接引用一個老實又經典的觀點。Richard Sutton 在 \u003Ca href=\"http:\u002F\u002Fwww.incompleteideas.net\u002FIncIdeas\u002FBitterLesson.html\" target=\"_blank\" rel=\"noopener\">The Bitter Lesson\u003C\u002Fa> 裡寫過一句話，現在還是很有殺傷力。\u003C\u002Fp>\u003Cblockquote>“The bitter lesson is that general methods that leverage computation are ultimately the most effective, and by a large margin.” — Richard Sutton\u003C\u002Fblockquote>\u003Cp>這句話的意思很直接。別太迷信手工巧思。真正能打的系統，常常是把算力和訓練方法吃滿。只是現在多了一層：不一定要把參數做得超大，才叫強。\u003C\u002Fp>\u003Ch2>Claude Opus 4.5 可能走哪條路\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Fclaude-4\" target=\"_blank\" rel=\"noopener\">Claude 4\u003C\u002Fa> 系列沒有公開參數數字。這很正常。前沿模型廠商現在幾乎都不愛講這件事。因為市場已經不太買單單純比大小了。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775207385570-sw8q.png\" alt=\"Claude Opus 4.5 和 GPT 到底多大\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>Anthropic 近年的重點很明顯。它很在意 coding、長上下文、工具使用，還有 agent 工作流。這些能力對產品比較實際。你拿來寫程式、整理文件、跑流程，才真的有感。\u003C\u002Fp>\u003Cp>所以如果 Claude Opus 4.5 或 4.6 的參數規模，落在 GPT-4o 類似區間，我一點也不意外。現在的競爭重點，早就不是誰喊出最大數字，而是誰用更少成本做出更好的體驗。\u003C\u002Fp>\u003Cul>\u003Cli>\u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002F\" target=\"_blank\" rel=\"noopener\">Anthropic\u003C\u002Fa> 沒公開 Claude Opus 的參數數字\u003C\u002Fli>\u003Cli>\u003Ca href=\"https:\u002F\u002Fopenai.com\u002F\" target=\"_blank\" rel=\"noopener\">OpenAI\u003C\u002Fa> 也沒公開 GPT-4 系列的完整大小\u003C\u002Fli>\u003Cli>GPT-4o 的 200B 到 300B，明顯低於 1T 級別想像\u003C\u002Fli>\u003Cli>較小部署腳印，通常更利於大規模上線\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>對開發者和買家，差在哪裡\u003C\u002Fh2>\u003Cp>如果你真的要選模型，別只盯著排行榜。很多產品根本不需要 1.6T 那種級別。200B 到 300B 的模型，只要速度夠快、價格合理、回答品質穩，就已經很夠用。\u003C\u002Fp>\u003Cp>這對 SaaS 團隊很重要。因為 \u003Ca href=\"\u002Fnews\u002Fai-coding-fast-trust-bottleneck-zh\">AI\u003C\u002Fa> 成本現在常常不是訓練，而是每天的推論帳單。只要使用量一上來，GPU 成本就會咬人。模型越能省，產品越容易活下來。\u003C\u002Fp>\u003Cp>還有一個現實問題。你要做的是客服、摘要、程式輔助，還是多步驟 agent？不同任務對模型的要求差很多。對很多場景來說，延遲低 30%，比參數多 3 倍更有感。\u003C\u002Fp>\u003Cul>\u003Cli>\u003Cstrong>訓練成本：\u003C\u002Fstrong>通常會隨模型規模快速上升\u003C\u002Fli>\u003Cli>\u003Cstrong>服務成本：\u003C\u002Fstrong>常是 AI 產品真正的瓶頸\u003C\u002Fli>\u003Cli>\u003Cstrong>延遲：\u003C\u002Fstrong>模型越精簡，通常回應越快\u003C\u002Fli>\u003Cli>\u003Cstrong>產品適配：\u003C\u002Fstrong>coding、摘要、agent 常看效率，不看面子\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>為什麼大家開始少談參數\u003C\u002Fh2>\u003Cp>這幾年，模型廠商越來越少公開參數數字。原因很簡單。第一，大家學會了參數不是全部。第二，公開太多，競爭對手也能更快推測架構方向。\u003C\u002Fp>\u003Cp>另外，市場也變成熟了。以前大家會拿 7B、13B、70B 互相比。現在更常問的是：上下文長度多少？工具調用穩不穩？coding 表現如何？價格每百萬 To\u003Ca href=\"\u002Fnews\u002Ftrivy-docker-images-fresh-supply-chain-attack-zh\">ke\u003C\u002Fa>n 幾美元？\u003C\u002Fp>\u003Cp>我覺得這是好事。因為這代表討論終於回到實用面。開發者要的不是一個好看的數字，而是一個能穩定跑在產品裡的模型。\u003C\u002Fp>\u003Ch2>所以 Claude Opus 4.5 到底多大\u003C\u002Fh2>\u003Cp>老實說，外部沒辦法精準知道。沒有官方數字，就只能看公開線索和產品表現。可是方向已經很清楚了：前沿模型正在往更高效率走。\u003C\u002Fp>\u003Cp>如果 Claude Opus 4.5 真的跟 GPT-4o 站在同一個量級，那它的重點就不是「多大」，而是「每個參數能做多少事」。我會押注這條路。接下來你該看的，也不是誰喊出更誇張的數字，而是誰能把成本壓低，還維持品質。\u003C\u002Fp>\u003Cp>對開發者來說，最實際的問題只有一個：你的產品，真的需要超大模型嗎？很多時候，答案其實是否定的。\u003C\u002Fp>","GPT-4 常被估到 1.6 兆參數，但 GPT-4o 可能只有 200B 到 300B。Claude Opus 4.5 的真實大小沒公開，重點其實是成本、延遲和效能比。","zhuanlan.zhihu.com","https:\u002F\u002Fzhuanlan.zhihu.com\u002Fp\u002F2020711987777159880",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775207388141-adee.png","research","zh","280d30d6-b080-4de0-b89b-fd239d8775fc",[17,18,19,20,21,22,23,24,25,26],"Claude Opus 4.5","GPT-4o","GPT-4","參數估計","大型語言模型","Anthropic","OpenAI","AI 成本","推論成本","台灣開發者",8,"2026-04-03T09:09:28.833454+00:00","2026-04-03T09:09:28.792+00:00",{"tags":31,"relatedLang":45,"relatedPosts":49},[32,34,36,37,39,41,42,44],{"name":17,"slug":33},"claude-opus-45",{"name":23,"slug":35},"openai",{"name":21,"slug":21},{"name":19,"slug":38},"gpt-4",{"name":18,"slug":40},"gpt-4o",{"name":25,"slug":25},{"name":22,"slug":43},"anthropic",{"name":26,"slug":26},{"id":15,"slug":46,"title":47,"language":48},"claude-opus-45-gpt-parameters-estimate-en","How Big Are Claude Opus 4.5 and GPT Models?","en",[50,56,62,68,74,80],{"id":51,"slug":52,"title":53,"cover_image":54,"image_url":54,"created_at":55,"category":13},"4fa896da-9616-425a-92bc-c1d7d5861ff9","streamma-multi-agent-reasoning-latency-zh","StreamMA 讓多代理推理邊想邊傳","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780554786134-1w1d.png","2026-06-04T06:32:32.769423+00:00",{"id":57,"slug":58,"title":59,"cover_image":60,"image_url":60,"created_at":61,"category":13},"f31f51ba-4445-4e43-9bda-31e70f53d42b","audio-language-models-arbitration-reversals-zh","音訊模型不是聽不懂","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780553877373-ux95.png","2026-06-04T06:17:27.890159+00:00",{"id":63,"slug":64,"title":65,"cover_image":66,"image_url":66,"created_at":67,"category":13},"447ac6c9-477b-45c8-bec2-ff94dc4cf5d4","stride-training-data-attribution-sparse-recovery-zh","STRIDE 讓訓練資料歸因快 13 倍","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780552979370-897a.png","2026-06-04T06:02:29.149166+00:00",{"id":69,"slug":70,"title":71,"cover_image":72,"image_url":72,"created_at":73,"category":13},"33c9a55c-a8c0-4367-b742-f4567d1e98e3","mathematicians-warn-ai-could-distort-math-zh","數學界警告 AI 會扭曲證明標準","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780504386035-080l.png","2026-06-03T16:32:29.415063+00:00",{"id":75,"slug":76,"title":77,"cover_image":78,"image_url":78,"created_at":79,"category":13},"5c3cb90f-7efd-426f-8c09-32a303f82be9","humanoid-gpt-zero-shot-motion-tracking-zh","Humanoid-GPT：用 GPT 擴大動作追蹤","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780469319284-znpc.png","2026-06-03T06:47:34.463464+00:00",{"id":81,"slug":82,"title":83,"cover_image":84,"image_url":84,"created_at":85,"category":13},"e3a4b0f7-03b3-43c6-ae51-906b337c5c2f","ipt-vlms-hidden-space-reasoning-zh","IPT 讓 VLM 更會想像隱藏空間","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780468394735-1k40.png","2026-06-03T06:32:46.560029+00:00",[87,92,97,102,107,112,117,122,127,132],{"id":88,"slug":89,"title":90,"created_at":91},"f18dbadb-8c59-4723-84a4-6ad22746c77a","deepmind-bets-on-continuous-learning-ai-2026-zh","DeepMind 押注 2026 連續學習 AI","2026-03-26T08:16:02.367355+00:00",{"id":93,"slug":94,"title":95,"created_at":96},"f4a106cb-02a6-4508-8f39-9720a0a93cee","ml-papers-of-the-week-github-research-desk-zh","每週 ML 論文清單，為何紅到 GitHub","2026-03-27T01:11:39.284175+00:00",{"id":98,"slug":99,"title":100,"created_at":101},"c4f807ca-4e5f-47f1-a48c-961cf3fc44dc","ai-ml-conferences-to-watch-in-2026-zh","2026 AI 研討會投稿時程整理","2026-03-27T01:51:53.874432+00:00",{"id":103,"slug":104,"title":105,"created_at":106},"cf046742-efb2-4753-aef9-caed5da5e32e","adaptive-block-scaled-data-types-zh","IF4：神經網路量化的聰明選擇","2026-03-31T06:00:36.990273+00:00",{"id":108,"slug":109,"title":110,"created_at":111},"53a0dc54-0371-4e40-8d5e-74e94a73840c","geometry-aware-similarity-metrics-for-neural-representations-zh","超越距離測量：用微分幾何重新理解神經網路","2026-03-31T06:01:01.241968+00:00",{"id":113,"slug":114,"title":115,"created_at":116},"fee7d472-a775-4b1d-bbc2-1e8bca1bbf8b","on-the-fly-repulsion-in-the-contextual-space-for-rich-divers-zh","讓AI繪圖更有創意：用排斥力提升生成多樣性","2026-03-31T06:01:25.439673+00:00",{"id":118,"slug":119,"title":120,"created_at":121},"a9901203-d69b-447b-8854-15d14eab32b4","vision-aided-beam-prediction-cnn-eca-zh","影像輔助波束預測升級 CNN","2026-04-01T10:00:25.8073+00:00",{"id":123,"slug":124,"title":125,"created_at":126},"b55e7dd4-0a24-4b3d-804d-b0309a03f498","triple-band-fss-mimo-antenna-sub-6-ghz-zh","三頻 FSS MIMO 天線瞄準 sub-6 GHz","2026-04-01T13:18:36.857305+00:00",{"id":128,"slug":129,"title":130,"created_at":131},"f68290bd-e7f3-4b30-ba22-dcd4e0130a66","openclaw-1299-repos-eight-weeks-analysis-zh","OpenClaw 1299 個 Repo 的資料解讀","2026-04-02T05:03:45.208411+00:00",{"id":133,"slug":134,"title":135,"created_at":136},"ed9f80eb-eb02-4d35-8ad4-0ddf428751dd","beam-coherence-aware-combining-mmwave-mimo-zh","毫米波 MIMO 的雙階合併法","2026-04-02T05:27:26.897188+00:00"]