WWDC 2026: iOS 27 and the multi-provider AI Extensions system — Claude, Gemini and ChatGPT in Siri
🔎 Siri no longer belongs to a single provider
On June 8, 2026, Apple kicked off WWDC 2026 with an announcement that redefines its AI strategy since the founding of Apple Intelligence. Siri is no longer an assistant locked to a proprietary model or an exclusive partner. iOS 27 introduces an AI Extensions system that allows developers to plug Claude, ChatGPT, Gemini, or any other compatible LLM directly into Siri's voice interface, Writing Tools, and Image Playground.
This is a paradigm shift. Until now, Apple had signed an apparent exclusivity partnership with OpenAI for ChatGPT in Siri. Now, Apple's voice layer becomes a neutral hub where the user chooses their AI provider. For iOS developers, this means completely rethinking the way AI capabilities are integrated into an app.
Under the hood, Siri is rebuilt on a Gemini model with 1.2 billion parameters, according to developer documents shared by LushBinary. But this default engine can be replaced or complemented by third-party extensions. Control returns to the end user.
The essentials
- Siri rebuilt on Gemini 1.2T params: the default engine of iOS 27 is a massive Google Gemini model, with context awareness, screen perception and cross-app task execution capabilities detailed by 36Kr.
- Multi-provider AI extensions: developers can plug Claude, ChatGPT, Gemini or other models into Siri via a new extensions framework, according to Tom's Guide.
- SiriKit deprecated, App Intents mandatory: ByteIota and TechTimes confirm that SiriKit is formally abandoned. App Intents is the only path to Siri integration under iOS 27.
- homeOS previewed: a new OS dedicated to Apple home devices, based on the same AI stack as iOS 27, according to ABHS.
- Staggered availability: iGeeksBlog specifies that the developer beta has been available since June 8, with a progressive public rollout — English as a priority, China excluded, EU staggered.
Recommended tools
| Tool | Main use | Price (June 2026, check on site) | Ideal for |
|---|---|---|---|
| Xcode 27 | iOS 27 dev, on-device AI, App Intents | Free (included in Apple Developer Program) | iOS developers migrating to AI Extensions |
| Claude Opus 4.7 (Adaptive) | Agentic model for integration into Extensions | Varies by Anthropic plan | Complex tasks, multi-step reasoning |
| GPT-5.5 | High-performance generalist LLM | Varies by OpenAI plan | General use cases, existing OpenAI ecosystem |
| Gemini 3.1 Pro | Default Siri engine under iOS 27 | Varies by Google Cloud plan | Native integration, consistency with the Siri engine |
Siri rebuilt: what actually changes under iOS 27
Siri under iOS 27 is not a cosmetic update. Apple rebuilt the assistant from scratch around three architectural pillars: screen perception, cross-app task execution, and context awareness.
The underlying model is a 1.2-trillion-parameter Gemini, making it by far the most massive on-device model deployed on a smartphone to date. This choice of Google as the default engine confirms Apple's pragmatic approach: rather than building a competitive proprietary model, Apple buys the best available capability and differentiates on the system integration layer.
What matters for developers is that this default engine is replaceable. The user can go into settings and select another provider via AI Extensions. Real-world tests published by MapleFeather show that switching between Gemini, Claude, and ChatGPT as the default assistant works in the early betas, with notable differences in latency and response quality depending on the tasks.
What this means for model benchmarking
With Claude Opus 4.7 (Adaptive) scoring 94.3 in agentic and GPT-5.5 at 98.2, the models available via Extensions cover a broad spectrum. The choice of default model in Siri is therefore not just an interface preference — it is a cognitive capability choice. A user switching from Gemini 3.1 Pro (90 in agentic) to GPT-5.5 gains nearly 9 points of performance on complex reasoning tasks.
The AI Extensions System: The Game-Changing API
The central mechanism of this WWDC is the AI Extensions framework. It works like existing iOS extensions (SHARE, ACTION), but for language models. A developer creates an extension that exposes an LLM to the system. Siri, Writing Tools, and Image Playground can then consume this model seamlessly.
According to The Verge, this system is designed to be vendor-agnostic. Apple does not favor one partner. The end user chooses in the settings which model powers which feature. For AI publishers like Anthropic, Google, and OpenAI, this turns the iPhone into a direct distribution channel.
The strategic analysis published by Philipp Lohmar on Medium sums up the stakes well: Apple controls the hardware, the OS, and the voice layer. For the first time, the AI distribution layer on the iPhone becomes truly competitive as an open market. AI providers are no longer fighting solely on the web — they are fighting in the pockets of 1.5 billion iOS users.
How an AI Extension Works in Practice
An AI Extension is declared via a new category in the host app's Info.plist file. It exposes capabilities — text generation, summarization, image analysis, tool execution — that the system can route to Siri or Writing Tools. The framework manages the lifecycle, response caching, and token management.
Apple requires the extension to comply with a standardized communication protocol. Your extension cannot arbitrarily modify Siri's behavior — it receives a structured request and must return a structured response. It is a strict interface contract, designed to prevent drift and guarantee the consistency of the user experience.
Mandatory App Intents: the end of SiriKit
This is arguably the most structuring announcement for existing developers. SiriKit is formally deprecated as of iOS 27. According to ByteIota, there is no hard removal date, but SiriKit will no longer receive updates and new Siri features (AI Extensions, screen perception, cross-app tasks) will only be accessible via App Intents.
TechTimes estimates that developers have 2 to 3 years to complete the migration before SiriKit stops working. That is a generous timeframe on paper, but given the complexity of some legacy SiriKit integrations, you need to start now.
Why App Intents is superior
App Intents, introduced with iOS 16, offers a declarative architecture that is much more flexible than SiriKit. Instead of defining rigid intents with constrained vocabularies, App Intents allows you to describe actions as pure Swift entities, with typed parameters, injectable dependencies, and automatic discovery capabilities by the system.
Under iOS 27, App Intents gains new capabilities: direct integration with the AI Extensions pipeline, support for agentic shortcuts (Siri can chain several App Intents in sequence to accomplish a complex task), and background execution with shared context between apps.
Migration checklist for developers
- Audit your existing SiriKit integration: list all the INxxx intents used, custom vocabularies, and interaction flows.
- Map to App Intents: each
INIntentmust find anAppIntentequivalent. Simple intents (send a message, start a task) migrate easily. Intents with complex custom vocabularies require refactoring. - Implement
AppShortcutsProvider: this is the discovery mechanism that allows Siri to suggest your actions in the interface. - Test on the iOS 27 beta: the developer beta has been available since June 8, 2026. Install it on a secondary device and validate that your App Intents work with the new Siri engine.
- Consider an AI Extension: if your app has text generation or reasoning needs, evaluate whether an AI Extension adds value compared to Siri's default engine.
Claude, Gemini, ChatGPT: who wins in this new landscape?
The Extensions system significantly reshuffles the deck. Here is where the three main providers stand at the time of WWDC 2026, based on available benchmarks.
Comparison of models available via AI Extensions
| Model | Provider | Agentic Score | General Score | Main strength under iOS 27 |
|---|---|---|---|---|
| GPT-5.5 | OpenAI | 98.2 | 91 | Best agentic score, mature OpenAI ecosystem |
| Claude Opus 4.7 (Adaptive) | Anthropic | 94.3 | 90 | Adaptive mode, safety, long reasoning |
| Gemini 3.1 Pro | 87.3 | 92 | Default Siri engine, native integration | |
| Claude Sonnet 4.6 | Anthropic | 81.4 | 83 | Quality/latency ratio, less token-hungry |
| GPT-5.4 | OpenAI | 87.6 | 89 | Good performance/cost trade-off |
OpenAI emerges stronger in the agentic domain. GPT-5.5 dominates the benchmarks with 98.2, and the tool ecosystem around the OpenAI API remains the most mature. But OpenAI is losing its "privileged" partner status with Siri — it is now just one provider among others.
Anthropic has a major card to play. Claude Opus 4.7 in adaptive mode is designed to adjust its reasoning level based on the complexity of the task, which is ideal on mobile where resources and user patience are limited. The fact that Claude is already being tested as the default assistant in iOS 27, as shown by MapleFeather's tests, indicates that Anthropic was involved early in the Extensions program.
Google occupies an ambiguous but powerful position. Gemini 3.1 Pro is the default engine for Siri — that's a huge win in terms of query volume. But at the same time, Google is competing against its own models via Extensions. A user who installs the ChatGPT or Claude extension bypasses Gemini. Google is betting that the quality of the native integration will make up for it.
For developers hesitating between these models for their own integrations, our Google Gemini vs ChatGPT vs Claude comparison details the strengths and weaknesses of each provider according to use cases.
Xcode 27 and on-device AI: the new dev paradigm
TechRepublic lists Xcode 27 among the ten major announcements of this WWDC, and for good reason. The IDE now integrates an on-device AI layer that changes the development workflow.
Xcode 27 can run AI models directly on the developer's machine, without any network calls. This covers code completion, unit test generation, performance analysis, and — crucial for this WWDC — the generation of App Intents and AI Extensions skeletons. You describe your intent in natural language within a panel, and Xcode generates the corresponding Swift code with the types, parameters, and entities.
This on-device integration is strategic for companies developing on confidential projects that cannot send code to cloud APIs. Apple positions Xcode 27 as the first "privacy-first" IDE with built-in AI.
The impact on the development cycle
For a project migrating from SiriKit to App Intents, Xcode 27 significantly reduces the boilerplate work. The automatic generation of AppShortcutsProvider, the suggestion of EntityQuery for discovery, and the static validation of AI Extension contracts eliminate a large part of the manual work. The business logic and testing remain, but the scaffolding is automated.
homeOS : Siri IA takes over the living room
Less publicized than AI Extensions but potentially just as impactful: homeOS. ABHS reports that Apple previewed a new OS dedicated to home devices — HomePod, Apple TV, and potentially a future hardware hub.
homeOS shares the same AI stack as iOS 27: same default Gemini engine, same AI Extensions framework, same App Intents. This means that an AI Extension developed for iOS 27 theoretically works on homeOS without modification. For smart home developers, this is an opportunity to create specialized voice assistants that run on Apple's home hub.
Context is key here. An AI assistant in the living room doesn't have the same latency or privacy constraints as an assistant on a smartphone. homeOS could support heavier models or longer tasks, with the advantage of an always-listening microphone and a rich home context (who is in the room, which device is active, what the habits are).
Availability and regional constraints
Not all of these features will be available everywhere at the same time. iGeeksBlog details a gradual rollout that reflects the regulatory constraints Apple has been familiar with since the launch of Apple Intelligence.
The developer beta has been available since June 8, 2026, for Apple Developer Program members. The public rollout will follow this fall, with a schedule that prioritizes English (United States, United Kingdom, Australia, Canada) in the first wave. The European Union is subject to a staggered rollout due to the DMA (Digital Markets Act). China is excluded from the initial launch.
For European developers, this means that AI Extensions must be tested in English first, with careful attention to DMA compliance — particularly the obligation for Apple not to favor an AI provider in its selection interface. The Extensions system, by its agnostic nature, seems designed to anticipate this regulatory requirement.
Strategic impact: why Apple made this choice
Beyond the technical aspect, the decision to open Siri up to third-party models reflects a clear-eyed analysis on Apple's part. The company cannot win the foundation model war. OpenAI, Google, and Anthropic are investing tens of billions per year in compute and R&D. Apple cannot keep up this pace without sacrificing its margins.
What Apple can do, however, is control the distribution layer. The iPhone is the most widely used device in the world. If every AI model has to go through Siri to reach the mobile user, Apple once again becomes the gatekeeper — not of content, but of AI access. This is exactly the analysis that Lohmar develops: the AI distribution layer on the iPhone becomes competitive for the first time.
For AI providers, the implications are major. Being available as an AI Extension in iOS 27 is no longer a nice-to-have — it is a business necessity. Recent new AI tools that do not offer an iOS 27 extension will be disadvantaged compared to those that do. The AI market is fragmenting into two: web models (accessible via browser) and integrated models (accessible via OS). Apple is betting that the second channel will become dominant on mobile.
❌ Common mistakes
Mistake 1: thinking SiriKit still works under iOS 27
SiriKit is deprecated. It will continue to work for existing apps for 2 to 3 years according to TechTimes, but no new Siri capabilities (AI Extensions, screen perception, cross-app) will be accessible via SiriKit. Migrating to App Intents is not optional if you want to leverage the new features.
Mistake 2: building an AI Extension without studying the interface contract
The AI Extensions framework imposes a structured communication protocol. You cannot do just anything with the user's request. A poorly designed extension that returns responses that do not conform to the expected schema will be rejected by the system, or even automatically disabled. Read the Apple documentation on AIExtensionCapability before coding.
Mistake 3: ignoring on-device vs cloud latency
AI Extensions can run on-device or via cloud. But the user expects a voice response in under 2 seconds. If your extension calls GPT-5.5 in the cloud with a poor network, the experience will be degraded compared to the default Gemini engine which is optimized for on-device. Consider implementing a fallback and choosing the model suited to your latency constraint.
Mistake 4: not preparing DMA compliance for the EU
If you publish an AI Extension and your app is available in the EU, the DMA applies. Apple will have to prove that its AI assistant selection system is neutral. Your extension cannot be disadvantaged compared to those of "privileged" partners. Test from the beta that your extension appears in the selection list without discrimination.
❓ Frequently Asked Questions
Can I use DeepSeek or Grok as an AI Extension in Siri?
The Extension framework is open to any provider that implements the Apple protocol. If DeepSeek or xAI release a compliant extension, they will work. The best LLMs for coding include DeepSeek V4 Pro and Grok 4.1, which have competitive scores. Nothing technically excludes them; it all depends on whether each publisher wants to release an iOS 27 extension.
Does the AI Extension replace direct SDK integration in my app?
No, these are two complementary mechanisms. The AI Extension allows your model to be consumed by Siri and the system Writing Tools. Direct SDK integration remains relevant for in-app features that do not go through Siri. For AI tools for marketing or SEO, you can have both a system extension and an in-app integration.
Do I have to pay for Siri's default Gemini model?
No. The Gemini engine integrated into Siri is included in iOS 27 at no additional cost to the end user. It is Apple that pays Google for the requests. However, if you develop an AI Extension that calls a cloud model (GPT-5.5, Claude Opus 4.7), the API costs are your responsibility or that of your extension's user.
Do AI Extensions work on older iPhones?
No. AI Extensions and the new rebuilt Siri require the A18 chip or later. iPhone 16 and earlier models will not benefit from the new system. This is a hardware limitation linked to the NPU capabilities required for the 1.2T parameter Gemini model on-device.
Does homeOS replace the Home app?
homeOS is a distinct OS, not just an update to the Home app. The Home app will continue to exist on iOS, but homeOS runs on dedicated devices (HomePod, Apple TV, future hub). Developers who create AI Extensions for homeOS are targeting a fixed device with different constraints than a smartphone.
✅ Conclusion
iOS 27 marks the transition of Siri from a proprietary assistant to a neutral AI hub. The multi-provider Extensions system, combined with the deprecation of SiriKit and the requirement for App Intents, redefines iOS development for the years to come. Developers who migrate quickly will have a competitive advantage in distribution via Siri. To explore the available models and their integration, check out our guide to the best free LLMs and our full analysis of Siri Extensions in iOS 27.