iOS 27 Siri評測：程式碼證實可換Claude、ChatGPT | iOS 27 Siri Review: Code Shows Claude, ChatGPT Swap

9月 14, 2026

By Kit 小克 | AI Tool Observer | 2026-09-15

🇹🇼 iOS 27 Siri評測：程式碼證實可換Claude、ChatGPT

Siri評測這次要看的不是新功能，而是一份程式碼考古發現：開發者從iOS 27與macOS Golden Gate的私有框架挖出證據，顯示Apple把Siri設計成可以整顆AI大腦抽換的架構——不只原本傳聞的ChatGPT，連Claude都能接進去驅動Siri。但目前這扇門還沒打開，一般使用者現在完全碰不到。

iOS 27程式碼裡藏了什麼？

爆料者「pdfu」在私有框架裡找到兩套機制。第一套叫Model Delegation，讓Claude像現有內建的ChatGPT擴充功能一樣，以「Siri擴充」的身分出現在系統裡。第二套叫Inference Provider Protocol，直接把Apple伺服器端的Siri模型整個換掉，pdfu實測時甚至用GPT-5.6 Terra跑通了流程。

在Inference Provider機制下，外部模型會拿到Siri的規劃提示詞（planner prompt）和完整工具定義，可以請求執行系統動作、存取個人資料，最後結果再由Siri用自己的介面和語音講給你聽——換句話說，使用者感覺不到背後換了引擎。

Apple為什麼要幫Siri留這條後路？

Siri重建計畫已經延宕超過一年，去年WWDC釋出的Apple Intelligence也被普遍認為追不上前沿模型。與其賭自家模型能追上Claude或GPT系列，Apple似乎選擇先把介面和資料存取的架構搭好，讓未來不管哪家模型領先，都能直接插進Siri裡使用。這也呼應了業界另一個趨勢：把「前端體驗」和「背後大腦」拆開設計，方便隨時抽換。

現在能用嗎？隱私風險呢？

答案很明確：不能用。Apple還沒把Model Delegation的授權開放給第三方，這功能也完全不會出現在使用者介面上，目前純粹是工程佈局。但這不代表沒有值得留意的地方：

優點：架構已經打通，一旦開放，使用者理論上能自選Siri背後的模型
隱私疑慮：Inference Provider模式讓外部模型直接碰到系統動作權限與個人資料，開放前需要更嚴謹的權限拆分
現況：僅存在於程式碼與私有框架，沒有時間表，也沒有對外申請管道

對一般使用者來說，這篇Siri評測的結論很簡單：值得追蹤，但別急著期待下一版iOS就能把Siri換成Claude。

好不好用，試了才知道。

🇺🇸 iOS 27 Siri Review: Code Shows Claude, ChatGPT Swap

This Siri review isn't about a new feature — it's about a code archaeology find. Developers digging through iOS 27 and macOS Golden Gate's private frameworks found evidence that Apple built Siri to have a fully swappable AI backend, and not just for the already-rumored ChatGPT integration — Claude can plug in too. The door isn't open yet, though: none of this is available to regular users today.

What's Actually in the iOS 27 Code?

Code sleuth "pdfu" found two separate mechanisms in Apple's private frameworks. The first, called Model Delegation, lets Claude show up as a Siri extension the same way the existing built-in ChatGPT extension does. The second, the Inference Provider Protocol, replaces Apple's server-side Siri model outright — pdfu demonstrated it working end-to-end with GPT-5.6 Terra.

Under the Inference Provider Protocol, the outside model receives Siri's planner prompt and full tool definitions, can request system actions, and can reach personal data — then hands results back for Siri to present in its own interface and voice. Users would never see the swap happening.

Why Would Apple Build an Exit Ramp for Its Own Siri?

Siri's rebuild has slipped for over a year, and last year's Apple Intelligence rollout is widely seen as falling behind frontier models. Rather than betting everything on catching up to Claude or the GPT lineup, Apple appears to be building the interface and data-access plumbing first, so whichever model leads later can simply be plugged into Siri. It's the same pattern showing up elsewhere in the industry: decouple the front-end experience from the model doing the thinking.

Can You Use It Today? What About Privacy?

No — not yet. Apple hasn't granted the Model Delegation entitlement to any third party, and none of this is user-facing. It's purely engineering groundwork right now. Still, a few things are worth flagging:

Upside: the plumbing already exists — once opened up, users could theoretically pick which model powers Siri
Privacy concern: the Inference Provider path hands an outside model direct access to system actions and personal data, which will need much tighter permission scoping before any public release
Reality check: this only exists in code and private frameworks — no timeline, no sign-up, nothing to try

The takeaway on this Siri review: worth watching, but don't expect the next iOS update to actually let you swap in Claude.

Good or not, you won't know until you try it yourself.

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

搜尋此網誌

AI小貼士