Apple lets users pick third‑party AI

Apple’s AI problem is not really raw model quality. It’s control. Apple Intelligence has been built as a mostly Apple-shaped layer, with one notable escape hatch for ChatGPT. Now that may be widening. A Bloomberg-backed report says Apple is preparing an “Extensions” system in iOS 27, iPadOS 27, and macOS 27 that would let users pick third-party AI models for Apple Intelligence features like Siri, Writing Tools, and Image Playground. ### Wait — is this official yet? No. That matters. As of Monday, May 11, 2026, Apple’s public developer docs and newsroom pages talk about Apple’s own on-device model, Private Cloud Compute, Shortcuts integration, and app-level access through the Foundation Models framework. They do not announce a user-facing picker for outside AI models inside Apple Intelligence. So the news here is a report about Apple’s plans, not a shipped Apple feature. (macrumors.com) ### What would actually change? Right now, Apple Intelligence mostly feels like one system with one fallback. Apple already lets ChatGPT step in for some requests. The reported next step is broader: users would be able to set a preferred outside model as the default for Apple Intelligence features, instead of being limited to Apple’s built-in models plus OpenAI as a special case. The report names Gemini and Claude as examples. (developer.apple.com) ### What is this “Extensions” thing? Basically, Apple seems to be turning AI providers into plug-ins. The reported “Extensions” feature would let installed apps expose generative AI capabilities on demand inside Apple Intelligence. That means Siri, Writing Tools, and Image Playground could call into another company’s model the way iOS already calls into app extensions for keyboards, share sheets, or password managers. Apple would keep the interface. (macrumors.com) The model provider would supply the brains. ### Why would Apple do that? Because Apple’s current AI strategy has a gap. Apple has strong device integration and strong privacy positioning, but its own models have not clearly led the field on open-ended text or image generation. Letting users swap in Gemini, Claude, or others would patch that weakness without Apple giving up the front-end experience. It also fits Apple’s recent direction — the company already opened its on-device Foundation Models framework to developers in iOS 26, and added a “Use Model” action in Shortcuts that can work on-device or through Private Cloud Compute. (macrumors.com) ### Does this mean Apple is giving up on its own models? Not really. More like Apple is hedging. Apple’s docs still center the on-device model at the core of Apple Intelligence, and that model is what developers can tap for private, offline features. The reported change would sit above that layer. Apple still owns the OS, the permissions, the routing, and probably the approval rules. Users would get more choice, but inside Apple’s frame. (developer.apple.com) ### Why is privacy suddenly trickier? Because “Apple Intelligence” stops being one privacy story. Apple’s on-device model can work offline, and Apple has pushed that hard. Third-party models are different. Some may run mostly in the cloud. Some may keep prompts longer. Some may use data differently. If Apple opens the door, it also has to label very clearly which model is answering, where data goes, and what rules apply. The report even says Apple plans separate voices for Siri versus third-party AI responses, which hints at that boundary problem. (developer.apple.com) ### What changes for developers? Potentially a lot. Today, developers can already build on Apple’s on-device model for app features, and Apple pitches that as private, offline, and free of inference cost. A user-selectable third-party layer would create a second path — one where model vendors compete for default status and app presence inside Apple’s own assistant surfaces. That could turn Apple Intelligence into a distribution channel, not just a framework. (macrumors.com) ### So what’s the real significance? If this ships, Apple stops pretending one model stack can do everything well. It becomes the orchestrator instead. That is a bigger shift than it sounds. The winner would not just be the best model. It would be the model that can live inside Apple’s rules without making the whole experience feel less Apple. (macrumors.com) (apple.com)