跳到主要內容

Gemini 3.5 Pro三度延期:Google砍掉重練追GPT-5.6 | Gemini 3.5 Pro's Third Delay: Google Restarts From Scratch

By Kit 小克 | AI Tool Observer | 2026-07-30

🇹🇼 Gemini 3.5 Pro三度延期:Google砍掉重練追GPT-5.6

Gemini 3.5 Pro 又跳票了,這已經是第三次。Google 執行長 Sundar Pichai 在 5 月 19 日還公開承諾「一個月內」交付,結果 6 月的第一個目標沒趕上,7 月 17 日的第二個死線也沒守住,重建後的版本至今仍未正式發布。對一家把 AI 當作核心戰場的公司來說,這不是小事。

三次延期,一次比一次嚴重

第一次延期發生在 6 月,原因是 Gemini 3.5 Pro 的程式碼能力沒有達到 Google 內部設定的目標,團隊試著用更新訓練資料的方式補救,結果反而讓表現更差。第二次延期問題更深層:Google 發現模型在「遞迴工具呼叫」(recursive tool-calling)與 SVG 生成上有結構性缺陷,單靠微調根本修不好,DeepMind 決定直接砍掉原本的基礎模型,從 Gemini 3 的基礎重新預訓練。這已經不是修 bug,而是重蓋地基。

根據 Bloomberg 引述十位現任與前任 Google 員工的說法,問題核心在於原始架構本身就撐不起設計目標。重建後的版本雖然完成,但仍有幻覺與可靠性落差,程式碼表現也還沒回到內部基準線之上——這就是第三次延期的原因。

對手不等人:GPT-5.6 與 Claude 已經量產部署

Google 慢下來的同時,競爭對手沒有停。OpenAI 在 7 月 9 日推出 GPT-5.6(分為 Sol、Terra、Luna 三個版本),同時上線 ChatGPT Work 企業產品;Anthropic 的 Claude Fable 5 與 Sonnet 5 也早已進入企業客戶的正式生產環境。當對手的旗艦模型已經在處理真實工作流程,Google 的競爭窗口正在快速關閉。

補位方案:Gemini 3.6 Flash

為了不讓 API 生態系停擺,Google 已經註冊 Gemini 3.6 Flash 與 Gemini 3.5 Flash Light 這兩個名稱,做為過渡方案。要注意的是,這兩款並非更強的旗艦模型,只是速度更快、體積更輕的版本,目的是維持開發者活躍度,而不是解決 Gemini 3.5 Pro 本身的問題。

悄悄開始的下一步:Gemini 4 已在預訓練

更耐人尋味的是,Google 在同一份承認 3.5 Pro 持續延期的聲明中,幾乎是用一句話帶過地透露:Gemini 4 的預訓練已經開始。這個時間點被外界解讀為一種「轉移焦點」的公關操作——與其讓外界聚焦在延期本身,不如丟出「下一代已經在路上」的消息稀釋負面情緒。

對開發者與企業用戶來說,現階段最實際的策略是:別再等 Gemini 3.5 Pro,先用現有的 GPT-5.6 或 Claude 把工作流程跑起來。好不好用,試了才知道。


🇺🇸 Gemini 3.5 Pro's Third Delay: Google Restarts From Scratch

Gemini 3.5 Pro has missed its deadline for the third time. Google CEO Sundar Pichai publicly promised delivery "within a month" on May 19, yet the model blew past its June target, missed a second deadline on July 17, and the rebuilt version still hasn't shipped. For a company treating AI as its core battlefield, this is not a minor slip.

Three Delays, Each Worse Than the Last

The first delay came in June: Gemini 3.5 Pro's coding performance fell short of Google's internal targets, and an attempted fix — updating the training data — actually made results worse. The second delay revealed a deeper problem: Google discovered structural failures in recursive tool-calling and SVG generation that couldn't be patched through fine-tuning alone. DeepMind scrapped the original base model entirely and restarted pretraining from a Gemini 3 foundation — not a bug fix, but rebuilding the foundation from scratch.

According to Bloomberg, citing ten current and former Google employees, the root issue was that the original architecture simply couldn't hit its design goals. The rebuilt version, while complete, still shows hallucinations and reliability gaps, with coding performance still below internal benchmarks — hence the third delay.

Rivals Aren't Waiting: GPT-5.6 and Claude Are Already in Production

While Google slows down, competitors keep shipping. OpenAI launched GPT-5.6 (in Sol, Terra, and Luna variants) on July 9, alongside the new ChatGPT Work enterprise product. Anthropic's Claude Fable 5 and Sonnet 5 are already running in production for enterprise customers. As rivals' flagship models handle real workflows today, Google's competitive window is closing fast.

The Stopgap: Gemini 3.6 Flash

To keep its API ecosystem from stalling, Google has registered the names Gemini 3.6 Flash and Gemini 3.5 Flash Light as interim releases. Important caveat: these are not more capable flagship models — just faster, lighter versions meant to keep developers engaged, not a fix for Gemini 3.5 Pro itself.

Quietly Underway: Gemini 4 Pretraining Has Begun

More intriguing is that Google buried a one-line admission in the same statement acknowledging 3.5 Pro's continued delay: pretraining for Gemini 4 has already started. Observers read the timing as a deflection — rather than let the narrative sit on the delay itself, Google offered "the next generation is already coming" to dilute the bad news.

For developers and enterprises, the practical move right now is simple: stop waiting on Gemini 3.5 Pro and build your workflow on GPT-5.6 or Claude today. 好不好用,試了才知道 — you won't know until you try it.

Sources / 資料來源

延伸閱讀 / Related Articles


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

這個網誌中的熱門文章

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?