OpenAI Spud 完成訓練：GPT-6 幾週後上線？ | OpenAI Spud Training Done: Is GPT-6 Weeks Away?
By Kit 小克 | AI Tool Observer | 2026-03-30
🇹🇼 OpenAI Spud 完成訓練：GPT-6 幾週後上線？
OpenAI Spud 是 OpenAI 下一代旗艦語言模型的內部代號，已於 2026 年 3 月 24 日完成預訓練（Pre-training）。執行長 Sam Altman 在內部備忘錄中形容這個模型「非常強大，足以真正加速整體經濟」，並預計將在「幾週內」正式發布。它究竟是 GPT-5.5 還是 GPT-6？目前 OpenAI 尚未確認命名，但這或許是自 GPT-4 以來最值得關注的一次模型更新。
Stargate 超算中心：10 萬張 H100 訓練出的模型
根據 The Information 的報導，Spud 在德州 Abilene 的 Stargate 設施中完成訓練，動用超過 10 萬張 NVIDIA H100 GPU，是目前已知最大規模的單一模型訓練集群之一。為了釋出算力，OpenAI 甚至關閉了 Sora 影片生成服務——這個取捨清楚說明了公司的戰略優先序：語言模型才是核心賭注，影片生成可以等。
超級應用藍圖：ChatGPT + Codex + 瀏覽器合而為一
Spud 不只是更強的聊天機器人。OpenAI 計劃讓它成為「超級應用」（Super App）的核心引擎，將以下三個元件整合為單一介面：
- ChatGPT：對話與推理介面
- Codex：程式設計代理人（Coding Agent）
- Atlas：AI 原生瀏覽器
對開發者來說，這意味著未來一個 API 呼叫可能同時處理搜尋、推理與寫程式，大幅簡化現有 AI 工作流的整合複雜度。這也是對 Claude Code 與 Google Gemini 生態系的直接挑戰。
GPT-6 還是 GPT-5.5？命名不是重點
OpenAI 近年習慣保守命名（o3、o4-mini、GPT-5.4），因此即使 Spud 是實質意義上的 GPT-6，正式名稱可能也讓市場失望。但重要的不是數字——部分 OpenAI 員工描述 Spud 的能力「和我們見過的完全不同」，這才是真正值得關注的訊號。等 benchmarks 出來，再評斷這個形容詞是否名副其實。
數學家 Terence Tao：AI 已成可信賴的共同作者
就在 Spud 完成訓練的同一個月，菲爾茲獎得主、數學家 Terence Tao 在新發表的論文中，使用 ChatGPT Pro 協助完成一個關鍵推導：AI 識別出傅立葉分析中的對偶性技巧，Tao 驗證後正式採用為論文的一部分。這正好實現了他 2023 年的預言——「AI 將在 2026 年成為值得信任的數學共同作者」。這不是 AI 取代數學家，而是人機協作模式正在真正成熟。如果現行的 GPT-5.4 已能做到這件事，Spud 的能力邊界又在哪裡？
開發者應該做什麼準備？
- 關注 OpenAI API 更新日誌，Spud 上線後 context window 與 agent 功能可能大幅擴充
- 評估現有 RAG 流程是否需要因應更長的 context window 調整（GPT-5.4 已達 105 萬 token）
- 留意 Codex + ChatGPT 整合後的 新 agentic API 介面，可能改變現有 LangChain、AutoGen 工作流
- 不要過度依賴未發布功能規劃產品路線圖；等 API 文件公開再行動
Sam Altman 說「事情進展比我們預期的還快」。這次我選擇相信他。但好不好用，試了才知道。
🇺🇸 OpenAI Spud Training Done: Is GPT-6 Weeks Away?
OpenAI's next flagship model, codenamed Spud, completed pretraining on approximately March 24, 2026. In an internal memo reported by The Information, CEO Sam Altman called it a "very strong model" that can "really accelerate the economy" — and said it would ship within "a few weeks." The central question: is this GPT-5.5, or have we quietly arrived at GPT-6?
Built at Stargate: 100,000 H100 GPUs
Spud was trained at OpenAI's Stargate facility in Abilene, Texas, using over 100,000 NVIDIA H100 GPUs — one of the largest single training runs in AI history. To free up that compute, OpenAI shut down its Sora video generation service. The trade-off is revealing: language model supremacy is the core strategic bet, and video generation can wait.
The parameter count remains undisclosed. OpenAI employees have described Spud's capabilities as "very different from what we've seen before" — but no benchmark results have been released publicly. Speculation is running hot across developer forums and HackerNews.
The Super App Play: ChatGPT + Codex + Browser in One
Spud is not just a better chatbot — it is designed to power OpenAI's planned super app, unifying three separate products into a single interface:
- ChatGPT — conversational reasoning interface
- Codex — AI coding agent
- Atlas — AI-native browser
For developers, this could mean one API endpoint replacing three separate integrations. It is also a direct challenge to Claude Code's growing developer mindshare and Google's Gemini ecosystem.
GPT-6 or GPT-5.5? The Name Does Not Matter
OpenAI has not confirmed the model's name. Given the company's recent conservative versioning habits (o3, o4-mini, GPT-5.4), a splashy "GPT-6" rebrand is not guaranteed. The branding may feel anticlimactic even if the capability jump is real. Focus on the capability delta, not the version number. When insiders describe a model as "very different from what we've seen before," that is the signal worth tracking.
The Terence Tao Signal: AI as Trustworthy Co-Author
In March 2026, Fields Medal mathematician Terence Tao published a paper in which ChatGPT Pro contributed a key proof — identifying a Fourier duality technique in a sub-problem, which Tao then verified and formally adopted. He had predicted in 2023 that AI would become a "trustworthy mathematical co-author" by 2026. That prediction just came true, one month before Spud even launches.
This is not AI replacing mathematicians. It is AI becoming a genuine collaborative tool — in one of the hardest intellectual domains humans work in. If current GPT-5.4 can do this, what does Spud look like?
What Developers Should Watch
- API release notes: Spud will likely appear in OpenAI's API within days of public launch — watch for the new model ID in the docs
- Context window: GPT-5.4 already hit 1.05M tokens; Spud may extend this further
- Agentic primitives: The Codex + ChatGPT merger hints at new agent APIs that could reshape existing LangChain and AutoGen workflows
- Do not pre-plan around rumors: Wait for the API documentation, run your own evaluations, judge by your actual use case
Sam Altman said "things are moving faster than many of us expected." This time, I believe him. But you won't know what it can actually do until you try it. 好不好用，試了才知道。
Sources / 資料來源
- OpenAI CEO Sam Altman Teases 'Very Strong' Model — The Decoder
- Is This GPT-6? OpenAI Bets Everything on Spud — Trending Topics EU
- OpenAI Spud & Terence Tao Math Proof 2026 — Revolution in AI
延伸閱讀 / Related Articles
- Mistral Small 4 開源：三合一 MoE 模型，免費商用 | Mistral Small 4: Open-Source 119B MoE That Does It All
- MCP 安全漏洞大爆發：AI Agent 最危險的攻擊入口 | MCP Security Crisis: The Hidden Attack Surface in Your AI Stack
- Gemini 3.1 Flash Live：Google 即時語音 Agent API，開發者搶先試 | Gemini 3.1 Flash Live: Build Real-Time Voice AI Agents with Google
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言