Apple settled a $250 million lawsuit in May 2026 for overpromising on Siri’s AI capabilities during the iPhone 16 launch. Two years of broken promises, ending in the largest consumer AI settlement to date. On June 8, at Tim Cook’s final WWDC keynote as CEO, Apple showed what it actually built to replace the Siri that earned that lawsuit: a completely rebuilt assistant powered by a custom 1.2-trillion-parameter Google Gemini model, running on Apple’s own Private Cloud Compute servers.
Then Apple did something more consequential than the Siri rebuild itself. It announced Extensions, a framework in iOS 27 that lets every iPhone, iPad, and Mac user choose which AI model powers their device. Gemini is the default. But users can switch to Anthropic’s Claude or OpenAI’s ChatGPT with a single settings toggle.
Apple just turned 2.2 billion active devices into an AI marketplace. For every frontier lab in the world, the distribution implications are enormous.
The $1 Billion Gemini Deal
Apple’s negotiations to rebuild Siri played out across most of 2025 and into early 2026. The company evaluated three options: build a frontier model internally, license from Anthropic, or license from Google.
Apple chose Google. The deal, signed in January 2026, gives Apple a custom 1.2-trillion-parameter Gemini model at approximately $1 billion per year. The model uses a mixture-of-experts architecture that activates only a relevant subset of parameters per query, meaning the full 1.2 trillion parameters never fire simultaneously. Apple runs the model on its own Private Cloud Compute servers (not Google’s infrastructure), so efficiency directly impacts Apple’s hardware costs.
That custom model is roughly eight times larger than the largest cloud model Apple built internally. The company determined that closing the gap would require sustained capital and talent investment it preferred to redirect toward hardware and services.
Google already pays Apple approximately $20 billion per year for default search engine placement on iOS. The Gemini deal creates a second revenue stream flowing in the opposite direction: Apple paying Google for AI capabilities, while Google pays Apple for distribution. The net economics of the Apple-Google relationship just became considerably more complex.
Why Anthropic Lost the Siri Contract
Anthropic was Apple’s first choice. Claude’s safety reputation and strong performance on reasoning benchmarks made it the natural partner for a privacy-focused platform.
The negotiations collapsed over price. Anthropic wanted “several billion dollars a year” for the Siri contract, with terms that could have doubled the annual cost over three years. Apple’s counteroffer sat closer to the $1 billion annual figure it eventually agreed to with Google.
The gap was unbridgeable. Anthropic, valued at $965 billion and preparing for its own IPO, chose to hold firm on pricing. Apple walked.
This decision will be studied for years. Anthropic turned down what would have been the largest consumer distribution deal in AI history: default placement on every iPhone, iPad, and Mac. The company bet that its model quality would earn users regardless of default status. With the Extensions framework, Apple gave Anthropic a path to test that thesis without the exclusive contract.
What the Rebuilt Siri Actually Does
The rebuilt Siri is not an incremental upgrade. It is a new product.
For the first time, Siri has a standalone app resembling ChatGPT’s interface: a text input field, microphone toggle, conversation history, and the ability to attach images and files. Swiping down from the iPhone’s center Home Screen triggers a new “Search or Ask” feature in the Dynamic Island, with a glowing animation during processing and transparent cards displaying results.
The deeper change is context. Siri now accesses emails, messages, files, photos, calendar data, contacts, and notes. It sees what’s on screen and acts on it. An address that appears in a text message can be added to a contact card without manual entry. A photo can be edited and shared in a single conversation thread. These are the multi-step, cross-app tasks that the previous Siri could never execute, the exact capabilities Apple advertised during the iPhone 16 launch and then failed to deliver.
iOS 27 drops support for iPhone 11, iPhone 11 Pro, iPhone 11 Pro Max, and the second-generation iPhone SE. Every remaining supported device gets the rebuilt Siri. Apple is counting on that compatibility cut to accelerate an upgrade cycle into iPhone 17 hardware later this year.
Extensions: The iPhone as AI Marketplace
The Extensions framework, launching with iOS 27 this fall, represents Apple’s most significant platform policy decision since the App Store.
Users open Settings, navigate to Apple Intelligence and Siri, and choose their preferred AI model: Gemini (the default), Claude, or ChatGPT. The chosen model powers Writing Tools, Image Playground, and Siri’s cloud reasoning tasks. Each model retains its own distinct voice, so users always know which one responded.
Third-party AI providers don’t need an exclusive billion-dollar deal with Apple. They need an App Store app that supports the Extensions API. Apple collects its standard platform fees. The model providers compete on quality.
This reframes Apple’s entire AI strategy. The company is not trying to build the best model. It is positioning itself as the platform where the best models compete for 2.2 billion users. The App Store of AI.
What the Distribution Math Means for Every AI Lab
For Anthropic, the Extensions framework offers the biggest consumer expansion opportunity in the company’s history. Apple has approximately 2.2 billion active devices. Even 5% adoption would deliver 110 million new Claude users, more than double Anthropic’s current estimated user base. Claude’s web traffic already grew 306% in a single quarter, reaching 8.2% global market share against ChatGPT’s dominant 54.7%. Extensions gives Anthropic a channel to accelerate that trajectory without spending a dollar on consumer marketing.
The risk is the default effect. Google pays Apple $20 billion per year precisely because defaults are sticky. Most users never change them. If Gemini-powered Siri is “good enough,” the vast majority of iPhone users will never open Settings to switch.
For OpenAI, the dynamic has shifted. The company already had an iPhone presence through the existing ChatGPT integration in Apple Intelligence. Extensions formalizes that arrangement but also demotes it: OpenAI is now one of three options, competing on equal footing. For a company that just filed for a $1 trillion IPO and reports 400 million weekly active ChatGPT users, being one choice among three on someone else’s platform is an unfamiliar position. The broader question: as Apple, Google, Microsoft, and Amazon all become AI distribution platforms, does OpenAI need its own hardware to guarantee distribution?
For Google, WWDC 2026 delivered a rare double win. Gemini is the default AI on the world’s most popular phone, and Google still collects $20 billion per year in search placement fees. The company now holds the two most valuable default positions on iOS. Whether Apple’s board is comfortable with that level of Google dependency under a new CEO is a question John Ternus will face early in his tenure.
Cook’s Last Strategic Decision
Tim Cook announced on April 20 that he would step down as CEO on September 1, handing the role to hardware chief John Ternus. WWDC 2026 was Cook’s final keynote.
The decision he is leaving behind shapes Apple’s next decade. He chose to make Apple a distribution platform for other companies’ AI rather than a builder of frontier models. This is the same strategic logic that built the App Store: control the platform, let others compete on it, collect fees from every transaction.
The App Store worked because Apple controlled both the hardware and the operating system. With AI, the value layer sits in the model, not the device. If Gemini or Claude or ChatGPT becomes the reason people buy an iPhone, leverage shifts from Apple to the model provider. Cook is betting that won’t happen, that users buy iPhones for the hardware, the ecosystem, and the privacy guarantees, not for whichever AI model runs underneath Siri.
Ternus inherits that bet. His background in hardware engineering (he led the M-series chip transition and Vision Pro development) suggests Apple will continue investing in on-device AI silicon while outsourcing cloud intelligence. Whether that split holds as models become more capable will define his tenure.
Running enterprise infrastructure for two decades, I’ve watched this exact pattern play out in every major platform shift. Apple is doing what every smart IT operation does: buy commodity intelligence from whoever offers the best price-to-performance ratio, then invest in the integration layer that makes it useful. The Extensions framework is that integration layer. The model underneath is, increasingly, a commodity. Apple just made that argument to 2.2 billion consumers. The AI labs now have to prove it wrong.