AI
“Code sleuth @itspdfu found private frameworks in iOS 27/macOS 27 ('Golden Gate') revealing Apple engineered Siri so its AI model can be swapped for third-party models, Claude can appear as a Siri extension, and GPT-5.6 can fully replace Siri's server-side model, while Siri retains its voice and controls iPhone tasks like Reminders,…”
Plain restatementA developer using the handle pdfu examined private (non-public) frameworks in iOS 27 and macOS 27 "Golden Gate" and demonstrated two mechanisms: (1) a Model Delegation API in App Intents that lets Anthropic's Claude appear as a selectable Siri extension, and (2) an "Inference Providing" protocol in Model Manager Services that lets an outside model such as GPT-5.6 substitute for Apple's server-side Siri model while Apple's Siri layer retains the voice, interface, and execution of system actions.
Distortion codes this site does not recognise yet: capability_extrapolation, demo_to_product_conflation, unreleased_as_released. Not collectible until the field guide has an entry.
This post is largely accurate. A developer known as pdfu did find hidden, non-public code in iOS 27 and macOS 27 "Golden Gate" and posted videos on 13 September 2026 showing two things: Claude working as a selectable Siri extension, and OpenAI's GPT-5.6 Terra standing in for Apple's own server-side Siri model while Siri kept its voice, interface, and ability to carry out tasks such as setting reminders. MacRumors and other Apple outlets reported this on 14 September 2026, so the post's sourcing checks out. The important caveat, which the carousel does mention on one slide but the headline and caption do not, is that none of this is switched on. Both demonstrations required private Apple permissions that Apple has not given to any outside developer, and the Claude extension was a bridge the researcher built himself rather than anything Anthropic released. The only third-party option actually available is Apple's existing built-in ChatGPT extension. Apple has not confirmed that it intends to open Siri to other models, and the suggestion that EU rules are forcing its hand is presented by the press as a possibility, not a fact.
[drifted from the evidence:] Code sleuth @itspdfu found private frameworks in iOS [drifted from the evidence:] 27/macOS 27 ('Golden Gate') [drifted from the evidence:] revealing Apple engineered Siri so its AI model [drifted from the evidence:] can be swapped for third-party models, Claude [drifted from the evidence:] can appear as a Siri extension, and GPT-5.6 [drifted from the evidence:] can fully replace Siri's server-side model, while Siri retains [drifted from the evidence:] its voice [drifted from the evidence:] and controls iPhone tasks like Reminders, Messages and [drifted from the evidence:] Mail.
[added by the neutral restatement:] A developer using the handle pdfu examined private [added by the neutral restatement:] (non-public) frameworks in iOS 27 [added by the neutral restatement:] and macOS 27 "Golden Gate" [added by the neutral restatement:] and demonstrated two mechanisms: (1) a Model [added by the neutral restatement:] Delegation API in App Intents that lets Anthropic's Claude appear as a [added by the neutral restatement:] selectable Siri extension, and [added by the neutral restatement:] (2) an "Inference Providing" protocol in Model Manager Services that lets an outside model such as GPT-5.6 [added by the neutral restatement:] substitute for Apple's server-side [added by the neutral restatement:] Siri model while [added by the neutral restatement:] Apple's Siri [added by the neutral restatement:] layer retains [added by the neutral restatement:] the voice, [added by the neutral restatement:] interface, and [added by the neutral restatement:] execution of system actions.
Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.
The trace / claim to source
- A developer using the handle @itspdfu (also "mactracker" on MacRumors Forums) did examine private frameworks in iOS 27 and macOS 27 "Golden Gate" and published findings on 13 September 2026.
- The two named mechanisms are real as described: Model Delegation in App Intents, and an Inference Providing protocol in Model Manager Services.
- Claude was demonstrated appearing in Siri's "Ask…" menu as a selectable extension, and it produced a CSV file, which Siri cannot do natively.
- Claude relays system-touching tasks such as setting a Reminder back to Siri, which performs the action.
- GPT-5.6 Terra was demonstrated receiving Apple's Siri planner prompt and tool definitions and making tool calls that perform system actions, with the response returned through Siri's UI and voice.
- The claim that Siri keeps its voice is accurate for the inference-provider path, in pdfu's own words.
- MacRumors published this on 14 September 2026, so the post's source attribution and date are correct.
- The carousel's caveats are accurate: only the ChatGPT extension is live in the macOS 27 Golden Gate Release Candidate, and Claude is not enabled.
- The DMA framing matches MacRumors, which raises it as a possible influence rather than an established cause.
- Capability extrapolation: the headline claim says "revealing Apple engineered Siri so its AI model can be swapped for third-party models," stating Apple's design intent as established. MacRumors hedges twice, saying the architecture works with third-party models "at what appears to be a surprisingly deep level" and that it "shows how extensively Apple has engineered Siri for future model interoperability." The existence of private, unshipped hooks is consistent with several explanations, including internal evaluation of Apple's models against outside models. Apple has said nothing. The claim converts a plausible inference into a revealed intention.
- Demo to product conflation: the claim text as written says "Claude can appear as a Siri extension" and "GPT-5.6 can fully replace Siri's server-side model" with no mention that both required private entitlements Apple has not issued to anyone, and that the "Claude extension" is a bridge pdfu built himself around the Claude command-line tool rather than anything Anthropic shipped. The carousel does carry the availability caveat on slide 5, which substantially mitigates this, but the caption and the top-line claim do not.
- Exaggeration: the caption's "Apple's walled garden just cracked" describes an event that has not occurred. Nothing was opened. Unreleased private code was discovered by a researcher.
- Omitted qualifier: the "same Siri voice" detail belongs to the inference-provider path. In the Model Delegation path, AppleInsider reports that responses are shown prefixed with the word "Claude" while otherwise resembling the Siri AI interface, so the two paths are not interchangeable on this point.
- Apple's actual intent and plans. No Apple confirmation, documentation, or comment was found, and no timeline exists for opening either mechanism.
- Whether the DMA caused this architecture. MacRumors and others present it as a plausible pressure, not a finding.
- Independent verification of the demonstrations. Everything traces to one researcher's videos, which I could not view directly; I could read his written descriptions and the coverage of them. pdfu has a track record (he surfaced unannounced watchOS features previously), but no second party has reproduced this.
- Which specific Claude model was used, and whether "GPT-5.6" in the post means GPT-5.6 Terra specifically, as pdfu stated.
- Whether Messages and Mail specifically are controlled in the demos. Reminders and email search and summarization appear in the coverage; Messages is asserted by the post without a source I could confirm.
pdfu published two demonstrations on 13 September 2026. In his own words, "iOS 27 and macOS Golden Gate have private hooks that let apps add Siri Extensions and replace Siri AI's server backend with third-party models," with Claude running as a Siri Extension via the Model Delegation API in App Intents, appearing in Siri's "Ask…" menu and supporting the same flow as the built-in ChatGPT extension. He notes this "Requires the private com.apple.developer.model-delegation entitlement." For the second mechanism, pdfu writes that an app extension replaces Siri AI's server model with GPT-5.6 Terra using the Inference Providing protocol in Model Manager Services, with GPT-5.6 receiving Apple's native Siri planner prompt and tool definitions, making tool calls that perform system actions and receiving tool call results with personal data, and returning "a response through Siri AI's UI and voice, including interactive elements." He adds that requests can be seen logged on OpenAI's platform and that it uses a private ExtensionKit extension point, com.apple.modelmanager.inferenceprovider. MacRumors reported the same on 14 September: pdfu "has uncovered iOS 27 and macOS Golden Gate private frameworks that show Apple has designed its new Siri architecture to work with third-party AI models at what appears to be a surprisingly deep level," with Model Delegation allowing Claude to appear as a Siri extension in the same way as the existing built-in ChatGPT extension. In the demo, the user chooses Claude via an "Ask…" contextual menu and asks Siri to "Ask Claude" to set a reminder; Claude interprets the request and Siri creates the reminder, the implication being that if the request requires an Apple system feature, Claude hands the task back to Siri; in another example Claude receives a request to create a CSV file, something Siri itself cannot handle. On availability, MacRumors states: the "Ask…" implementation is currently limited to the ChatGPT extension in the macOS 27 Golden Gate Release Candidate, so Claude is not yet available, and "Apple has not yet opened up the model delegation entitlement to third parties and it isn't front-facing to users, but it at least shows how extensively Apple has engineered Siri for future model interoperability." iClarified is blunter: "Claude is not available as a user-facing option, and Apple has not opened the Model Delegation entitlement to third-party developers. For now, the frameworks reveal what Apple's new Siri architecture is capable of, not what the company will ultimately make available." 9to5Mac: "You cannot yet activate this, it's visible in the code but not currently enabled." On the regulatory framing: MacRumors says the EU's Digital Markets Act "may have helped shape Apple's approach here, as it requires Apple to give third parties effective access to iOS hardware and software features available to Apple's own services, and the European Commission has specifically said this principle extends to Siri." On how the Claude extension was actually built, pdfu's own earlier technical write-up explains that his Claude bridge implements Apple's internal AgentIntent, and because the extension is sandboxed it forwards each request over a framed localhost connection to a companion LaunchAgent, which invokes the authenticated claude CLI with streaming JSON output and a scoped set of MCP tools.
Complete reasoning
The reply is formatted for pasting into the thread where the claim is circulating.
Compact share page: ai.trueseeker.com/s/03a7b9636d0e/Wx2ixPNCOpSO82eL1FLDcOrZrCs
Ask this case
Answers come only from the case file above; nothing is added.
Who found this and what exactly did they find?
A developer using the handle @itspdfu (also known as mactracker on MacRumors Forums) examined private, non-public frameworks in iOS 27 and macOS 27 'Golden Gate' and published findings on 13 September 2026. He identified two real mechanisms: a Model Delegation API in App Intents and an Inference Providing protocol in Model Manager Services.
Can I actually use Claude or GPT-5.6 with Siri right now?
No. Both demonstrations required private Apple entitlements that have not been issued to any outside developer, and Claude is not enabled even in the current macOS 27 Golden Gate Release Candidate. The only third-party option actually available to users is Apple's existing built-in ChatGPT extension.
Was the Claude integration something Anthropic built?
No. The Claude extension was a bridge the researcher built himself, using Apple's internal AgentIntent framework to forward requests to a companion process running the authenticated Claude command-line tool. It was not released or provided by Anthropic.
Does this prove Apple intends to let other companies power Siri?
Not established. The case file shows Apple has built the technical hooks for this, but Apple has not confirmed any intention to open Siri to outside models. Reporting describes this as showing what the architecture is capable of, not what Apple will ultimately release.
Is it true that EU regulation is forcing Apple to do this?
This is not confirmed. MacRumors raises the EU's Digital Markets Act as something that may have shaped Apple's approach, since it requires interoperability with Siri, but this is presented as a possible influence, not an established fact.