On June 8, 2026, Apple finally shipped the Siri it promised two years ago — and it took a partnership with its biggest rival to get there. At the WWDC 2026 keynote, Apple announced Siri AI, an entirely new assistant rebuilt on a generation of Apple Foundation Models developed in collaboration with Google's Gemini — a deal both companies confirmed in a rare joint statement. For product and engineering teams, the headline isn't the demo. It's the architecture underneath it, and a set of developer platform changes that quietly reshape how apps plug into an AI assistant.
This post covers what was actually announced, how the Apple–Google arrangement works technically, the release timeline and hardware requirements, and what teams building on iOS — or building AI features anywhere — should take from it.
What Apple Announced at WWDC 2026
Siri AI is not a feature update; Apple describes it as an entirely new version of Siri, deeply integrated into iPhone, iPad, Mac, Apple Watch, and Vision Pro. The capabilities Apple demonstrated are the ones it first promised at WWDC 2024 and then delayed:
- Personal-context search. Siri AI can search across messages, emails, photos, and other personal data to answer questions grounded in your own information.
- On-screen awareness. The assistant can answer questions about whatever is currently on screen, or follow up with a web lookup.
- A conversational interface with memory. There is now a dedicated Siri app, and conversation history syncs privately across devices via iCloud.
- System-level presence. Siri AI lives in the Dynamic Island on iPhone, integrates with Spotlight on macOS, and extends to CarPlay and AirPods.
All of this arrives across iOS 27, iPadOS 27, macOS 27 (named "Golden Gate"), watchOS 27, and visionOS 27. Developer betas shipped on keynote day, public betas land in July via the Apple Beta Software Program, and the general release comes as a free update in fall 2026. Siri AI itself ships as a beta "later this year," English first, and access in the current developer betas sits behind a waitlist.
Inside the Apple–Google Gemini Partnership
The most consequential line of the keynote was an acknowledgment: Apple said it collaborated with Google and the Gemini family of models to build the third generation of Apple Foundation Models — the AFM 3 family that powers Apple Intelligence and Siri AI. The joint statement is unambiguous about the shape of the deal: a multi-year collaboration under which the next generation of Apple Foundation Models is based on Gemini models and Google cloud technology. Press reports put the arrangement at roughly $1 billion per year built around a custom ~1.2-trillion-parameter Gemini model, though those figures come from secondary reporting rather than either company.
What makes this more interesting than "Apple licensed Gemini" is how Apple threaded it. According to Apple's machine learning research post, AFM 3 is a family of five models "custom-built in collaboration with Google." AppleInsider's analysis captures the nuance: the shipped models were built with the aid of Gemini and its technology through distillation and training — not by embedding Google's models directly. Apple gets frontier-class capability; the weights it ships remain its own.
The architecture is tiered. Simple requests run on-device on Apple's models. Heavier reasoning routes through Private Cloud Compute to the top-tier model, AFM 3 Cloud Pro, which Apple says is comparable in quality to Gemini's frontier models — a vendor assertion with no independent benchmark yet. The infrastructure detail worth noticing: Apple worked with Google and Nvidia to extend Private Cloud Compute to Nvidia GPUs running in Google Cloud, while claiming the same verifiable privacy guarantees it built for its own silicon. Apple's privacy story is now a three-company story, and CNBC's coverage of the keynote made clear how central that framing was to the pitch.
What Developers Get: An Open Foundation Models Framework
For developers, the quieter announcements may matter more than Siri itself. Apple's Apple Intelligence guide for WWDC26 lays out two significant changes.
First, the Foundation Models framework is no longer Apple-models-only. Any model that conforms to a new public Swift LanguageModel protocol can plug in — Apple's on-device models, but also cloud models like Gemini and Claude, with official Swift packages from Google and Anthropic coming soon. In practice, Apple just shipped a provider-abstraction layer at the OS level: the same app code can swap between an on-device Apple model and a third-party cloud model behind one interface. That is the pattern we've argued teams should build for themselves in using AI features without making the product fragile — keeping the model behind a boundary you control — now formalized as a platform API.
Second, the economics changed for small teams. Developers enrolled in the App Store Small Business Program whose apps have fewer than 2 million total first-time downloads get access to the new Apple Foundation Models running on Private Cloud Compute at no cloud API cost. The free tier covers Apple's models only — third-party Gemini or Claude calls are still billed by their providers — and Apple has not disclosed pricing beyond the threshold. But for indie apps and early-stage products, server-grade model inference just became a free platform capability rather than a line item.
App Intents Is Now the Front Door
The second platform shift: App Intents schemas are now the primary way — with SiriKit deprecated, effectively the only way — to expose your app's content and actions to Siri AI's natural-language layer. Apple's pitch is that these schemas are structures Siri "understands deeply," built on years of model training, so adopting them makes app capabilities reachable through conversation with relatively little code.
Two supporting pieces shipped alongside:
- Onscreen-awareness APIs (session 343) let Siri resolve "this" and "that" against what the user is looking at —
NSUserActivityfor a single primary entity,appEntityIdentifierannotations for lists and collections, plus entity identifiers on notifications, Now Playing sessions, and alarms. - AppIntentsTesting, a new framework for validating intents through the same infrastructure Siri, Shortcuts, and Spotlight use — no UI automation required.
The testing framework deserves more attention than it got. When an assistant becomes a primary surface for your app, "does Siri correctly invoke our intent" becomes a regression class, and until now there was no sane way to put it in CI. This is the same discipline we describe in the role of evaluation in AI product quality: behavior that lives in a model's interpretation of your inputs needs its own automated checks, run where decisions happen. Apple just made that possible for intent resolution.
There's a design implication too. An app exposed through App Intents is an app whose features get invoked without its UI on screen — which forces clarity about what each action does, what it needs, and when it should confirm before acting. Teams that have already thought through when automation should ask for approval will find that work pays off directly here.
Devices, Dates, and the 12GB Divide
The compatibility floor for Apple Intelligence is broad: iPhone 15 Pro and later, M1-or-later iPads and Macs, the A17 Pro iPad mini, the MacBook Neo, Vision Pro, and recent Apple Watch models paired with a supported iPhone. iOS 27 itself still supports every iPhone back to the iPhone 11.
But the full Siri AI experience is hardware-gated. The most powerful on-device model — including expressive voices and advanced dictation — requires 12GB of unified memory. On iPhone, that means iPhone 17 Pro, iPhone 17 Pro Max, and iPhone Air; the base iPhone 17 at 8GB misses the cut, alongside iPad (M4 or later) and Mac (M3 or later) with 12GB. If you're building features that assume on-device Siri AI capabilities, plan for a split user base for several hardware generations: device capability tiers are now a product-design input, not just a performance footnote.
One caution on availability: regional rollout beyond "English first" is genuinely unclear. Claims that Siri AI will skip the EU or China did not survive verification in our research, but Apple has not published a definitive regional timeline either. Treat any confident statement about EU availability as speculation for now.
What Product Teams Should Take From This
Strip away the keynote polish and three durable lessons remain.
Pragmatism beat purity. Apple spent two years trying to get to frontier-class Siri on its own, then bought capability from a competitor and engineered the integration so its privacy architecture and model ownership survived. Most teams face a miniature version of this decision — build the model layer or buy it — and Apple's answer is instructive: buy the capability, own the interface, and invest your differentiation where users can feel it. The interface layer is where assistants succeed or fail, a point we've made in prompt interfaces need product thinking.
Provider abstraction is becoming table stakes. When the OS vendor ships a LanguageModel protocol so apps can swap Gemini for Claude for an on-device model without rewriting features, single-provider coupling in your own codebase starts to look like technical debt.
The assistant is a new distribution surface — with a deadline. Apps that adopt App Intents schemas are discoverable through natural language the day Siri AI reaches users this fall; apps that don't are invisible to it. The betas are out now, the testing framework exists, and the window between "early adopter" and "catching up" is one OS cycle.
Apple's bet is that users care less about whose model answers and more about whether the answer is private, fast, and woven into their device. Whether that holds is the most interesting product question of the fall. The engineering question is already settled: the assistant layer is now part of the platform, and it ships in September.
