Apple allows swappable AI models

Apple’s AI story may be changing from “here’s our model” to “pick the one you want.” That’s the real news here. The reported plan for iOS 27, iPadOS 27, and macOS 27 is an “Extensions” system that would let Apple Intelligence hand jobs to outside models like Google Gemini or Anthropic Claude instead of relying on one default assistant. (bloomberg.com) ### What is Apple actually changing? Right now, Apple Intelligence feels like a layer Apple controls, with ChatGPT bolted on in a limited way for certain requests. The reported shift is much broader. Users would be able to choose a preferred third-party model for system features across Apple’s software, not just a one-off handoff when Siri gets stuck. (bloomberg.com) ### Why does “Extensions” matter? Because it changes Apple’s role. Instead of trying to win every model race itself, Apple can become the platform that routes requests to the best available model. That is a very Apple move in one sense — control the interface, privacy prompts, settings, and device integration — but it is also unusually open for Apple, because the intelligence underneath could come from rivals. (bloomberg.com) ### Which models are in the picture? The names that keep showing up are Gemini and Claude. MacRumors says Apple has signed a deal with Google and plans to use a Gemini-based model for Apple Intelligence and Siri features in iOS 27, while Bloomberg’s reporting says users will be able to choose from multiple outside services. Claude is also named as an example option. (macrumors.com) ### Where would you feel this on the phone? In the obvious places first — Siri, Writing Tools, and Image Playground. Those are the parts of Apple Intelligence where model quality is easiest to notice. If one provider writes better summaries, another follows instructions better, and a third is stronger on images, model choice stops being a geeky setting and starts becoming a real product decision. (macrumors.com) ### Does this mean everything runs in the cloud? Not necessarily. And this is the catch. Apple’s current AI pitch depends heavily on a split between on-device processing, Apple’s own private cloud setup, and external providers when needed. A swappable-model system makes that routing more complicated. Some tasks may still stay local. Others could jump(macrumors.com)rs for latency, privacy, and cost. This last part is an inference from how Apple’s current architecture works and from the reported multi-model design. (theverge.com) ### Why would Apple do this now? Because Apple still looks behind in frontier models, but it has enormous distribution. If the company cannot outbuild OpenAI, Google, and Anthropic fast enough, it can still own the device layer where people actually use AI. Basically, Apple gets to sell the switchboard while others supply the brains. That is a much safer position than betting the whole platform on one in-house model suddenly catching up. (engadget.com) ### What does this do to Apple’s partners? It gives them access to Apple users, but under Apple’s rules. Providers would likely need compatibility through their apps or system hooks, and Apple would still control defaults, permissions, and interface design. So yes, Gemini or Claude could gain reach — but Apple would still own the customer relationship in the places that matter most. (letsdatascience.com) ### Bottom line If this ships, the important Apple AI product won’t just be Siri or a single model. It will be the chooser. Apple may be deciding that the winning move in AI is not to be the smartest model on the phone — it’s to be the operating system that decides which model answers. (bloomberg.com)