Skip to content

Apple Stopped Trying to Win the Model Race. At WWDC, It Made the Model a Setting.

DS
LDS Team
Let's Data Science
8 min
On June 8, Apple rebuilt Siri on a custom Google Gemini model it reportedly pays around a billion dollars a year to license, then shipped a Swift framework that lets any app swap between Apple's models, Gemini, and Claude without changing a line of code. Xcode 27 now routes coding work to Claude, Gemini, or OpenAI agents. The one place the new Siri will not ship is the European Union.

At 10 a.m. Pacific on Monday, Tim Cook walked onto the stage at Apple Park and, for the engineers watching, buried the most important sentence of the day in a product demo. Apple finally showed the rebuilt Siri it had been promising for two years. The detail worth stopping on was what sits underneath it: a custom version of Google's Gemini.

It was Cook's last Worldwide Developers Conference keynote as chief executive. He becomes executive chairman on September 1, handing the company to hardware chief John Ternus. For a CEO who spent a decade selling Apple silicon and on-device processing as the company's deepest advantage, the farewell came with an admission most observers underplayed.

One of the most valuable companies on earth has decided it will not build a frontier AI model. It will rent one, put it underneath its own assistant, and turn every rival's model into a component that developers select like a dependency. For anyone who builds software, that is the actual story of WWDC 2026.

Siri Now Thinks With Google's Model

The new assistant, which Apple is calling Siri AI, is a ground-up rebuild rather than another layer bolted onto the 2011 original. It holds a back-and-forth conversation, reads what is on your screen, and digs through your own messages, emails, and photos to answer a question. There is now a standalone Siri app that syncs conversation history across devices through iCloud. In a demo, it looked far more like ChatGPT or Claude than the Siri anyone has been yelling directions at for years.

The cloud intelligence behind it runs on what Apple described as Apple Foundation Models "custom-built in collaboration with Google and its Gemini models." Lighter requests stay on-device, on Apple silicon. The heavier reasoning routes to a Gemini-based model running on Apple's Private Cloud Compute servers, so Apple controls the infrastructure even though the model is Google's. Bloomberg reported before the keynote that the arrangement costs Apple roughly a billion dollars a year. Apple has not confirmed a figure.

That is a genuine philosophical reversal. Every other company that decided a frontier model was strategically essential built its own. OpenAI built GPT. Google built Gemini. Anthropic built Claude. Microsoft built MAI. Apple, sitting on one of the largest cash piles in corporate history, wrote a check to a competitor instead. Apple's framing on stage was that it is not trying to win the model race. It is trying to remain the device through which every model gets accessed.

For Developers, the Model Became a Line of Swift

The reversal goes deeper than Siri, and this is the part that matters for practitioners. Apple rebuilt its developer stack around the same idea: the model is now a swappable backend.

The Foundation Models framework is now a single native Swift API for talking to AI models. A new LanguageModel protocol lets an app point at Apple's own on-device models, Google's Gemini, or Anthropic's Claude, switching providers through Swift Package Manager with no changes to session code. The model an app uses becomes a configuration choice rather than a rewrite.

Xcode 27 carries the same logic into the editor itself, with a dual-engine design:

  • A local Neural Engine model handles real-time Swift suggestions on the developer's machine.
  • A cloud routing layer sends heavier analysis to coding agents from Anthropic's Claude, Google's Gemini, or OpenAI.
  • The agent can simulate an entire app, write and run tests, inspect visual changes through live previews, and drive the iOS Simulator through a new Device Hub.

Crucially for anyone already living inside the agent ecosystem, Xcode 27 speaks the open protocols the rest of the field uses. Developers can extend it with plug-ins, pull in external tools through the Model Context Protocol (MCP), and connect any agent that supports the Agent Client Protocol. GitHub and Figma are the first to ship one-tap integrations. Apple, the company most associated with the walled garden, just made its IDE model-agnostic and protocol-friendly.

LayerWhat powers itWho chooses
On-device Apple IntelligenceApple Foundation Models on Apple siliconApple
Siri AI cloud reasoningCustom Gemini model on Private Cloud ComputeApple
App features (Foundation Models API)Apple models, Gemini, or ClaudeThe developer
Xcode 27 coding agentClaude, Gemini, or OpenAI agentsThe developer
Apple Intelligence assistantChatGPT, Gemini, or ClaudeThe user

The Picker Was the Rumor. This Is the Product.

The consumer-facing version of all this is a settings menu. Across Apple Intelligence, users can choose which model answers, with options including ChatGPT, Gemini, and Claude, and Gemini set as the default. That ends the single-provider ChatGPT arrangement Apple shipped with iOS 18. The outline leaked weeks ago, and we covered it when it did in Apple Is Ending ChatGPT's iPhone Monopoly. The Gemini-under-Siri partnership itself traces back to the deal we wrote about in the Gemini era of Apple Intelligence. What shipped Monday is the confirmed, demoed product, and it is bigger than the rumor: not just a chatbot picker, but a model layer that runs from Siri down to Xcode.

There are limits worth knowing before the demos set expectations. Siri AI and the headline Apple Intelligence features require an iPhone 16 or later, or an iPhone 15 Pro. iOS 27 itself installs on phones back to the iPhone 11, so millions will get the update this fall without the AI that anchored the keynote. Siri AI itself slips to a beta later this year rather than launching with iOS 27. And images generated by Apple's tools now carry Google's SynthID provenance watermark, baking the Google relationship into the files themselves.

The Other Side: A Rented Brain and a European Wall

Two objections sit underneath the applause.

The first is dependency. Apple's privacy architecture is real, and running Gemini on Private Cloud Compute means Apple controls where data goes. But the company has now tied the quality of its flagship assistant to a competitor's model roadmap. Every time Gemini trails Claude or GPT on a benchmark that users feel, Apple feels it too, and it cannot simply ship a better model of its own to close the gap. Owning the device is a strong position. It is not the same as owning the intelligence.

The second is regulation, and it is not hypothetical. Apple confirmed that Siri AI will not ship on iPhone or iPad in the European Union with iOS 27 and iPadOS 27, and said it has no timeline to change that. Apple blames the Digital Markets Act directly, arguing that the EU's reading would force it to give any third-party assistant the same deep access to user data that Siri gets, which Apple says it cannot do safely. Apple proposed a workaround it calls a Trusted System Agent, plus an 18-month phased rollout. It says the European Commission rejected all of it. EU users will still get Siri AI on Mac, Apple Watch, and Vision Pro, just not on the two devices most people actually use. The features are also delayed in China while Apple works through local requirements. A model stack that Apple spent years and a billion dollars assembling can be switched off in an entire market by law.

The Bottom Line

The Bottom Line

Apple just told the industry where it thinks the value is, and it is not in the model. By renting Gemini for Siri and shipping a Swift API and an IDE that treat Claude, Gemini, and OpenAI as interchangeable backends, Apple is betting that owning the device and the distribution beats owning the intelligence. For developers, the model is now a setting. For Apple, the open question is whether the company that controls a billion-plus devices can stay indispensable once the thing those devices think with belongs to someone else.

If the bet is right, Apple becomes the front door to every frontier model and never has to win a benchmark. If it is wrong, it has handed the most important layer of its products to the one rival it cannot out-engineer, and discovered in Brussels that even a rented brain can be regulated out of the room.

Sources

Practice interview problems based on real data

1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.

Try 250 free problems
Free Career Roadmaps8 PATHS

Step-by-step roadmaps from zero to job-ready — curated courses, salary data, and the exact learning order that gets you hired.

Explore all career paths