Apple Stopped Pretending It Would Build the Model

For two years Apple's AI story rested on a single premise: that the assistant in your pocket would, eventually, be Apple's. The delays were framed as discipline. The privacy posture was framed as a moat. The missing features were framed as worth the wait. On Monday, June 8, in Tim Cook's final WWDC keynote before John Ternus takes the CEO chair on September 1, Apple quietly abandoned the premise. The new Siri runs on Google.

Not "calls Google for hard questions." Runs on it. The rebuilt assistant — rebranded Siri AI — is powered by a custom 1.2-trillion-parameter Gemini model that Apple licensed from Google under a deal first struck in January 2026 and confirmed on stage this week. The recurring figure across outlets is roughly $1 billion a year. The model runs on Apple's own servers under its Private Cloud Compute framing, but make no mistake about what changed: the intelligence answering your questions is Google's, rented.

The tell is the Extensions menu

The headline is the Gemini deal. The more revealing detail is the structure Apple built around it. iOS 27, iPadOS 27, and macOS 27 ship with an Extensions system that lets users choose which model handles Apple Intelligence requests — Google Gemini as the default, with ChatGPT and Anthropic's Claude as selectable options, each given a "distinct voice so users know which one responded."

Read that as an admission. A company confident in its own model does not build a clean abstraction layer that makes swapping in three competitors a settings toggle. Apple did exactly that — and in doing so turned the model itself into a commodity slot, a thing the user picks like a default browser. The framing being floated, that outsourcing the model is somehow Apple's "ultimate moat," is the kind of analysis that arrives after a decision is already made. The honest version is simpler: Apple could not ship a frontier assistant on its own models in a competitive window, so it bought one and built a polite menu around the fact.

That is not necessarily a bad decision. It may be the correct one. Siri has been the most-mocked product at the most-admired company in tech for the better part of a decade, and the gap to GPT-class assistants was not closing on Apple's roadmap. Licensing a 1.2-trillion-parameter model and wrapping it in Apple's interface, on-device personal context, and Dynamic Island integration gets a usable assistant into a billion pockets faster than any internal sprint could. The new Siri does the things people actually wanted: web search, image generation, summarization, coding help, file analysis, multi-step commands. The chatbot Apple users were quietly opening ChatGPT for is now the system assistant.

Who actually won

Google. Unambiguously. Analysts are already calling it "the most profitable enterprise AI contract in history," and the description holds up. Google collects roughly a billion dollars a year, gets its model embedded as the default intelligence layer on the highest-margin hardware platform on Earth, and does it while its existing search-default payments to Apple sit under antitrust scrutiny. Gemini was already shipping fast — Gemini 3.5 Flash landed this month with a 55 on the Intelligence Index, and 3.5 Pro is expected mid-June. Now it has the one distribution channel money normally can't buy: the position of Apple's first-party assistant.

For Anthropic and OpenAI, the Extensions slot is a smaller prize but a real one. This is the first time Claude is a selectable option on the iPhone at the system level, riding alongside ChatGPT. Anthropic spent the spring building enterprise reach — a Partner Hub, a $100M partner program, Claude Opus 4.8 landing in Microsoft Foundry and Excel's agent mode — and its web traffic reportedly jumped triple digits quarter over quarter. Being a one-tap default option inside Siri extends that surface to consumers. But "selectable option" is a worse position than "default," and default is Google's.

The strategic cost Apple isn't pricing in

There's a reason vertically integrated companies fight to own their core technology, and Apple wrote the book on it: control the whole stack and you control the experience, the margins, and your own destiny. Apple owns its silicon, its OS, its app distribution, increasingly its modems. The model is now the one layer of the most important computing platform of the decade that it rents from a direct competitor — one it is simultaneously litigating against over search payments and competing with in phones, maps, browsers, and ads.

Dependencies like this have leverage embedded in them. Pricing can move. Model access can be deprioritized. The roadmap for the assistant on a billion Apple devices is now partly Google's roadmap. Apple has clearly thought about this — the Extensions architecture exists precisely so it can swap providers, and a custom model running on Apple servers is less exposed than a raw API call. But optionality is not ownership. The ability to switch from Google to OpenAI is not the same as needing neither.

The most quietly significant line of the keynote wasn't about Siri at all. It was Cook handing the company to Ternus in September. The incoming CEO inherits an Apple that just made the largest concession of its AI era — that the company which insists on building everything itself decided the one thing it most needed to build, it would buy instead. Whether that reads in five years as pragmatism or as the moment Apple ceded the future of its assistant will depend entirely on what it does with the time the deal bought. For now, the most valuable real estate in consumer computing — the assistant on the iPhone — answers in Google's voice. Apple just made it official.