Apple is opening Siri. Siri Extensions, shipping with iOS 27 later this year, will let users route queries to third-party AI chatbots — Claude, Gemini, Grok, and others — directly from the Siri prompt and the Settings app. The change ends OpenAI’s exclusive arrangement on the Apple platform and reframes Siri from a single-vendor assistant into a configurable chatbot router. Pair this with the parallel news that Apple’s next-generation Foundation Models will be based on Google’s Gemini family (with on-device and Private Cloud Compute processing preserved), and the picture for May 2026 is unmistakable: Apple just made its strongest move yet to compete in the AI assistant category — by letting other AI companies do the heavy lifting where Apple’s own models fall short.
What’s actually new
The headline change is the Extensions system itself. In iOS 26 and earlier, Apple Intelligence shipped with a single integrated path to OpenAI’s ChatGPT for queries the on-device model couldn’t handle. The arrangement was exclusive. Users could disable it, but they couldn’t substitute a different chatbot. iOS 27 replaces that integration with an open Extensions framework: any AI provider can ship an Extension package, the user installs and authorizes it through Settings, and Siri will route queries to the chosen Extension when appropriate.
The pre-announced launch lineup includes Anthropic (Claude), Google (Gemini), xAI (Grok), and Perplexity. OpenAI’s ChatGPT remains an option but is no longer the default. The user picks one or more Extensions and configures Siri’s routing preferences — “ask before routing,” “always use Claude for these query types,” “use Gemini for searches with current information needs.” Multiple Extensions can be installed simultaneously; Siri’s routing logic decides per-query which Extension fits, or asks the user when there’s ambiguity.
The second major change is Apple’s foundation-model strategy. Apple’s prior Foundation Models — a ~3B parameter on-device model and a larger server-based model running on Apple Silicon Private Cloud Compute infrastructure — are being replaced over the next 12 months with models derived from Google Gemini, fine-tuned and adapted for Apple’s privacy and on-device constraints. Apple’s distinctive privacy architecture (Private Cloud Compute, on-device processing for most queries) is preserved; the model weights underneath shift from Apple-trained to Gemini-derived. The justification, per Apple’s announcements: capability gap. Apple’s internal models lagged the frontier; partnering with a frontier lab closes that gap faster than retraining from scratch.
Third, Apple is opening up the on-device Foundation Models framework to third-party developers. Where previously only a small set of system features used Apple Intelligence directly, third-party app developers can now invoke the on-device models through a public Foundation Models framework, with the same privacy guarantees Apple’s first-party features enjoy. Developers building apps that benefit from on-device AI — drafting, summarizing, translation, classification — get a privacy-respecting model essentially for free, without bundling their own.
The combined effect: Apple positions itself as the privacy-respecting AI platform layer, with multiple frontier-lab chatbots accessible through it, frontier model capability via the Gemini partnership, and on-device intelligence available to any developer who wants it. The platform play is more ambitious than anything Apple has previously committed to in AI.
Why it matters
- End of OpenAI exclusivity reshapes the AI distribution landscape. ChatGPT was the most-distributed third-party AI experience on consumer devices through OpenAI’s Apple deal. Losing exclusivity (without losing presence) tightens the competitive landscape immediately. Anthropic, Google, xAI, and Perplexity all gain access to ~1 billion iPhone users overnight.
- Apple becomes a routing layer, not a model provider. Siri’s new role is “pick the right AI for the job,” not “be the smartest AI.” This is the same pattern that worked for Apple in maps (offer Apple Maps, integrate Google Maps via app linking), browsers (Safari plus default-browser choice), and search (default-deal with Google). It’s the safer business position; the trade-off is ceding model competition to others.
- Gemini under the hood changes the AI infrastructure map. Apple shipping Gemini-derived foundation models means Google has effective distribution into one billion devices outside its own ecosystem. The competitive implication for OpenAI and Anthropic is meaningful — both wanted that distribution position; both lost it.
- Privacy posture remains Apple’s differentiator. Apple’s Private Cloud Compute architecture — encrypted, attested, no Apple-side data storage — is genuinely stronger than what cloud-AI providers offer directly. Routing through Apple to Claude or Gemini provides users a privacy guarantee that going to claude.ai or gemini.google.com doesn’t. This is a real product advantage.
- On-device Foundation Models for third-party developers unlock new app categories. Apps that previously couldn’t ship AI features because of cost (per-token API bills), privacy (sensitive data couldn’t go to a cloud API), or latency (cloud round-trip too slow) gain a free, private, fast on-device model. Expect a wave of new on-device AI features in third-party apps over the next year.
- The competitive pressure on Siri’s quality drops. If Siri can route hard queries to Claude or Gemini, Siri itself doesn’t need to be a frontier-model competitor. This reduces Apple’s R&D pressure and lets Apple focus on the product layer (UX, integrations, privacy) where it traditionally wins.
How to use it today
Siri Extensions ship with iOS 27 (expected fall 2026). Some pieces of the broader Apple Intelligence stack are usable today; others require waiting. Here’s the practical roadmap.
- Available now: the existing Apple Intelligence features in iOS 26 — Writing Tools, Image Playground, Smart Replies, Notification Summaries, on-device transcription. These are baseline capabilities that won’t change at iOS 27 launch beyond gradual quality improvements as the underlying models update.
- Available now for developers: the existing Apple Foundation Models framework. Third-party apps can already invoke the on-device 3B-parameter model for tasks like summarization, drafting, and classification.
// Swift — invoke the on-device Apple foundation model import FoundationModels let session = LanguageModelSession() let response = try await session.respond(to: """ Summarize this in two sentences: \(longText) """) print(response.content) - Available now via TestFlight beta: developers in the Apple Developer Program can sign up for the iOS 27 beta starting at WWDC 2026 (June). Test devices get early access to the Siri Extensions framework and the new Foundation Models APIs. If you’re building an app that benefits from these capabilities, get on the beta early.
- Wait for iOS 27 (consumer): general availability is fall 2026. At launch, the Settings app gains a new “Apple Intelligence → Extensions” section where users install and configure third-party AI chatbots. The first Extensions (Claude, Gemini, Grok, Perplexity) ship alongside iOS 27.
- For app developers integrating Extensions: the SiriKit Extensions API (subset of the broader Extensions framework) lets your app participate in the Siri routing layer. If your app provides an AI capability, you can ship an Extension that lets Siri route relevant queries to your app.
// Swift — minimal Siri Extension manifest (sketch) import AppIntents import SiriExtensions struct MyChatbotExtension: SiriChatbotExtension { static var name = "My Chatbot" static var description = "Answers questions using my company's domain expertise." static var supportedQueryTypes: [QueryType] = [ .general, .technicalSupport, .productInquiry, ] func handleQuery(_ query: SiriQuery) async throws -> ChatbotResponse { let answer = try await myBackend.processQuery(query.text, context: query.context) return ChatbotResponse(text: answer, citations: [], confidence: 0.85) } } - For enterprises planning the iOS 27 transition: review your MDM policies. Some organizations will want to allow some Extensions (Claude for staff using corporate accounts) and block others (consumer chatbots that might exfiltrate data). Apple’s MDM framework supports per-Extension policy; your IT team should plan the policy before users start asking.
How it compares
| AI assistant | Distribution channel on iPhone (post-iOS 27) | Privacy posture via Apple | Pricing for end users | Strength |
|---|---|---|---|---|
| Apple Siri (default) | Built-in | Strongest — on-device + Private Cloud Compute | Free | System integration, calendar/email/messages |
| Anthropic Claude (Extension) | Via Siri routing | Strong — Apple’s privacy proxy applies | Free tier + Claude Pro $20/mo | Reasoning, coding, careful tone |
| Google Gemini (Extension) | Via Siri routing + standalone app | Strong — Apple’s privacy proxy applies | Free tier + Pro $20/mo | Real-time info, multimodal, ARC-AGI-2 reasoning leader |
| xAI Grok (Extension) | Via Siri routing | Strong — Apple’s privacy proxy applies | Free tier + SuperGrok subscription | Less filtered tone, X integration |
| Perplexity (Extension) | Via Siri routing | Strong — Apple’s privacy proxy applies | Free tier + Pro $20/mo | Search-augmented answers with citations |
| OpenAI ChatGPT (Extension) | Via Siri routing (no longer exclusive) | Strong — Apple’s privacy proxy applies | Free tier + Plus $20/mo | Broad capability, large ecosystem |
The notable feature of the table: privacy posture is the same across all the third-party Extensions. Apple’s privacy-proxy architecture treats every Extension equivalently. Users don’t have to evaluate each AI provider’s individual privacy policy; they get Apple’s privacy guarantees regardless of which Extension answers the query.
The strategic implication: AI chatbot competition on iPhone moves to product capability, not data-handling posture. The user picks the chatbot that’s smartest for their query types; Apple handles the privacy infrastructure underneath. This is structurally similar to how the iOS App Store works for apps generally — Apple handles the platform-level guarantees; apps differentiate on what they do.
What’s next
Three trajectories worth watching for the rest of 2026 and into 2027.
Extension ecosystem expansion. The launch lineup (Claude, Gemini, Grok, Perplexity, ChatGPT) will expand. Expect Mistral, DeepSeek, regional players (Baidu’s ERNIE, Yandex’s YandexGPT), and vertical specialists (legal, medical, financial chatbots) to ship Extensions over the following 12 months. The Extensions framework is open to any provider that meets Apple’s review criteria.
Routing intelligence improvement. Siri’s logic for picking which Extension handles which query will iterate. The launch behavior will be conservative (“ask the user when uncertain”); over time, the routing will learn user preferences and route more queries automatically. The user-experience competition shifts from “is Siri smart enough” to “does Siri pick the right helper for the job.” This is a different problem with a different competitive dynamic.
The Gemini partnership will reshape Apple’s AI strategy. Adopting Gemini as the foundation-model base is a pragmatic short-term decision (faster path to frontier capability) with long-term implications. Apple will likely continue training its own Foundation Models in parallel and may swap back if its internal models close the capability gap. Or the Gemini partnership might deepen — Apple’s distribution plus Google’s models is a strong combination either company would want to preserve. Watch for follow-up announcements through 2027.
The deeper question: does Apple’s “platform layer for AI” strategy actually win? The historical pattern suggests yes — Apple consistently profits more from being the hardware-and-OS substrate than from competing in the application categories that run on top. The risk is that AI is sufficiently strategic that Apple needs to win the model layer too, and that ceding it to partners weakens the long-term position. Both readings are credible. The answer will become clear over the next 24-36 months as the AI ecosystem either consolidates around platforms (Apple’s bet) or fragments such that platform position matters less than model quality (the contrarian bet).
Frequently Asked Questions
When does iOS 27 ship?
Apple has confirmed a fall 2026 general availability, with the developer beta starting at WWDC in June 2026 and the public beta starting in July or August. Specific dates haven’t been announced yet; Apple typically announces the dot-zero release date at the September fall event.
Will Siri Extensions work on older iPhones?
The Foundation Models framework requires iPhone 15 Pro or newer (Apple Intelligence-capable devices). Siri Extensions specifically — the routing layer to third-party chatbots — will likely require the same hardware baseline, since the on-device model needs to evaluate which Extension to route to. Older iPhones running iOS 27 may get a degraded experience that routes everything to a cloud-based default rather than running the smarter routing on-device.
Do third-party Extensions get access to my Apple data?
No, not by default. Siri’s privacy architecture treats user data as confidential. When Siri routes a query to an Extension, only the query text (plus optional user-approved context like the active app or selected text) goes through. The Extension doesn’t get access to your contacts, messages, calendar, or other Apple data unless you explicitly grant per-query permission.
How does Apple’s foundation model partnership with Google affect on-device privacy?
The on-device model continues to run entirely on the device — your queries to it never leave your phone. The cloud model running on Private Cloud Compute infrastructure stays on Apple Silicon servers in Apple’s data centers, with the same encrypted, attested, no-storage architecture. The Gemini partnership changes which model weights run; it doesn’t change where the inference happens or who has visibility.
Should developers switch from cloud AI APIs to on-device Foundation Models?
Depends on the use case. On-device models are smaller (3B parameters) and meaningfully less capable than frontier cloud models on hard tasks. For drafting, summarization, classification, and similar bounded tasks, on-device is excellent — fast, private, free per call. For complex reasoning or domain expertise, cloud APIs still win. Most apps will use both: on-device for the easy things, cloud for the hard ones.
Will OpenAI’s position be hurt long-term by losing Apple exclusivity?
Probably some, but not catastrophically. ChatGPT remains the most-distributed AI product overall, with strong direct distribution via web and apps. Losing Apple’s default position pulls some marginal users to other Extensions; it doesn’t move ChatGPT’s existing power-user base. The bigger long-term question for OpenAI is whether Anthropic, Google, and xAI can win user preference on iPhone over time — if they do, OpenAI loses iPhone-originated growth even if its existing users stick.