Apple’s AI Reboot
The new Siri runs on a custom Gemini model costing Apple roughly a billion dollars per year.
Apple announced Siri AI at WWDC 2026 on June 8, a ground-up rebuild of its voice assistant whose most demanding cloud queries run on Google’s Gemini. Gemini handles the cloud tier while Apple retains the interface, app access, and privacy layer, giving each company distinct control over different parts of the same assistant.
The announcement came two years after Apple unveiled Apple Intelligence at WWDC 2024 with commitments to a substantially upgraded Siri that preceded more than a year of deferrals and left the promised features undelivered. A $250 million class-action settlement with iPhone buyers over those unfulfilled promises received preliminary court approval on May 5, five weeks before the WWDC 2026 keynote.
Two years overdue
Apple introduced Apple Intelligence at WWDC 2024 with a commitment to a substantially upgraded Siri, then repeatedly delayed the promised features with each subsequent software release. The centerpiece of that commitment was a Siri capable of understanding personal context and acting across apps, capabilities that Apple tied to the iPhone 15 Pro and the iPhone 16 lineup. Those delays stretched more than a year, as competing AI assistants continued to advance.
A class-action lawsuit filed by iPhone buyers alleged that Apple had advertised, for the iPhone 15 Pro and iPhone 16 lineup, Siri features that never shipped. The settlement totaled $250 million, covered purchases made between June 10, 2024, and March 29, 2025, and received preliminary court approval on May 5, 2026. Apple denied wrongdoing and said it had resolved the matter “to stay focused on doing what we do best.”
The Siri AI announcement at WWDC 2026 came two years after Apple first unveiled those Siri commitments and five weeks after the settlement received preliminary court approval. Daniel Newman, chief executive of the Futurum Group, described the announcement as a “prove-it moment” for Apple, arguing that the company’s record of delays had given him little reason to trust its new timeline.
A billion-dollar backend
Apple and Google announced a multi-year collaboration on January 12, 2026, confirming that Gemini models would power future Apple Intelligence features, including the rebuilt Siri. Apple’s June 8 Siri AI press release credits “the next generation of Apple Foundation Models” without mentioning Google or Gemini by name. Bloomberg’s Mark Gurman reported in November 2025 that Apple would pay Google roughly $1 billion per year for a custom Gemini model with approximately 1.2 trillion parameters, terms that neither company has publicly confirmed.
Siri AI routes queries across three infrastructure tiers, directing each request to a different layer based on its computational demands. Simple tasks remain on the device, processed by Apple’s own smaller neural models, while moderately complex requests travel to Private Cloud Compute, where Apple Silicon servers handle them in stateless, sealed nodes. The heaviest queries reach Google Cloud, where they run on Nvidia Blackwell B200 GPUs using a custom Gemini model with 1.2 trillion parameters and a mixture-of-experts architecture that activates only a relevant subset of those parameters for any given request. That custom model carries approximately eight times the parameter count of the 150-billion-parameter models that Apple had previously run in its own cloud infrastructure.
Before any query reaches Google’s infrastructure, Apple anonymizes it, strips Apple ID linkage, and tokenizes the data. A contractual provision bars Google from using those query streams to train future models. Google’s branding appears nowhere in the Siri AI experience, since Apple owns the interface, app access, and privacy layer throughout.
Fully embedded
Siri AI ships as a dedicated standalone app with conversation history that persists across devices through iCloud. Through personal context understanding, Siri searches across a user’s messages, emails, and photos to answer questions or take action across apps. That personal context extends to communication style, with Siri drawing on past exchanges in Mail and Messages to learn each contact’s expected register and punctuation.
Onscreen awareness allows Siri to read whatever appears on screen and answer questions about it across any app. Writing Tools can generate drafts, refine existing text, and proofread automatically as a user types, including within most third-party apps. For a user who typically sends a manager short bullet points, Siri AI populates new email drafts in Mail with that same format.
On iPhone, Camera Siri mode lets a user point at any subject and tap the shutter for an answer or action, including splitting a restaurant bill through Apple Cash or identifying a meal’s nutritional content. Visual Intelligence, previously limited to iPhone, now extends to iPad, Mac, and Apple Vision Pro. iOS 27 allows a user to designate a third-party AI service as the default Apple Intelligence provider, with Apple naming Claude, ChatGPT, and the Gemini app as options.
The fine print
Siri AI runs on iPhone 15 Pro and later, iPad and Mac with M1 or later, Apple Vision Pro, and Apple Watch Series 10 and later at launch. A developer beta launched on June 8, public betas follow in July, and the consumer release arrives alongside iOS 27 in September. iOS 27 installs on iPhone 11 and newer, but Siri AI itself requires more capable hardware, meaning that older iPhone owners receive the operating system without the assistant’s AI features.
Siri AI’s most advanced on-device model, which enables expressive voices and advanced dictation, requires at least 12GB of unified memory. Devices that meet that threshold include iPhone Air, iPhone 17 Pro, and iPhone 17 Pro Max; iPad with M4 or later; Mac with M3 or later; and Apple Vision Pro with M5. That threshold applies to a subset of compatible devices, so most Siri AI users receive a capable on-device model but not the most advanced one.
For EU users, the Digital Markets Act blocks Siri AI on iPhone, iPad, and Apple Watch at launch, with no timeline announced for those platforms. Apple proposed a mechanism that it called the ‘Trusted System Agent’ to let rival assistants access Siri AI’s device features in the EU and offered to phase in the arrangement over 18 months, but the European Commission rejected both. EU users who have a Mac or Apple Vision Pro receive Siri AI at launch on those platforms, making the restriction platform-specific within the region. In China, Siri AI remains unavailable while Apple works through local regulatory requirements, with no timeline announced.
Strings attached
At roughly $1 billion per year in Bloomberg’s accounting, the Gemini supply arrangement mirrors the Google Search default on Apple devices, giving Google a second recurring channel through Apple’s user base. Rebecca Haw Allensworth, an antitrust professor at Vanderbilt University, has argued that the Gemini deal “essentially creates a second exclusive pipeline,” applying the same distribution logic that animates the Justice Department’s ongoing antitrust appeal over Google’s search monopoly.
The September release carries the credibility deficit forward into the launch. Newman’s “prove-it moment” framing locates the test in adoption rather than capability, asking whether integration this deep can translate into daily use from a user base that watched two years of Apple’s AI commitments fail to materialize.
Apple frames Siri AI as a wide-device-lineup feature. The DMA’s exclusion of iPhone, iPad, and Apple Watch in the EU, the complete absence in China, and a 12GB memory threshold that restricts the most advanced on-device model to a hardware subset all define the actual launch reach more narrowly than Apple’s own framing suggests.