On the afternoon of June 9, inside Apple Park's Steve Jobs Theater, Craig Federighi gave the WWDC 2026 keynote and spent a surprisingly small fraction of it talking about hardware. The announcement that will matter most over the next several years was not a chip, not a display, and not a redesigned application. It was a settings screen.
Apple introduced a framework called AI Extensions — shipping with iOS 27, iPadOS 27, and macOS 27 — which allows users to designate a third-party AI model as the brain behind Siri, Writing Tools, Image Playground, and the rest of the Apple Intelligence feature set. Claude, Gemini, and ChatGPT are all in scope. Users can set one model as their default across the entire OS, or route different task types to different providers: Gemini for research, Claude for coding assistance, ChatGPT for creative writing. TechCrunch's full WWDC roundup confirmed the announcement alongside iOS 27's other major features. An earlier 9to5Mac report from May described the framework under the internal name "Extensions" before the keynote.
The surface story is choice. The actual story is distribution.
Apple has 1.4 billion active devices worldwide. Getting a model into a user's daily workflow — not as a separate app they have to consciously open, but as the thing that answers when they ask Siri a question, or activates when they tap Writing Tools in their messages — is a distribution problem that no AI lab has fully solved on its own. Anthropic has Claude.ai. Google has Gemini. OpenAI has ChatGPT. None of those are where most people spend most of their time on their phones. Siri is.
The Extensions framework changes this at the system level. A user who sets Claude as their default model does not need to open a separate app. Every voice query, every Writing Tools invocation, every Apple Intelligence feature call routes to Claude. This is the distribution reach that AI labs have been attempting through partnerships, developer deals, and aggressive App Store promotion — and Apple handed it to them through a settings page while retaining control of the frame.
That last clause is the one worth sitting with. Apple is not stepping aside. The Extensions framework routes queries to third-party models, but Apple controls the interface, the privacy handoff (what data gets sent to the model and under what terms), the voice selection, and the settings page where users make the choice. Apple's own Apple Intelligence features — on-device processing, Private Cloud Compute — continue as a first-party layer beneath the Extensions framework. A user running Claude through Siri in iOS 27 is getting Claude's capabilities through Apple's pipes, subject to Apple's policies, with Apple owning the user experience.
For the AI labs, the business implications are genuinely mixed. A model that becomes the preferred AI for millions of iOS users has distribution that no ad campaign can replicate. But a model running inside Extensions has less visibility into its own usage than a model running in a native app — the user experience belongs to Apple, and the relationship the lab builds with the user is mediated by a company with its own interests in that relationship.
The picture is further complicated by an announcement that received less coverage: Apple said at WWDC that developers with fewer than two million first-time App Store downloads would receive free access to Apple Foundation Models running on Private Cloud Compute. That is Apple's own models, offered at no cost to small developers, available alongside the Extensions framework that routes to third-party models. Apple is simultaneously opening the door to third-party AI and lowering the cost of its own AI to zero for the apps that can't afford to pay.
For the broader question of who controls the agentic web — a question we examined when Google proposed WebMCP — Apple's move adds a new layer. The competition for where AI runs is no longer only between model providers. The operating system is now a routing layer, and the companies that own operating systems have views about what routes through them.
For developers building iOS applications, the Extensions announcement creates a specific category of engineering uncertainty. If your app relies on a particular model's behavior — a generation style, a specific approach to function calling, a consistent output format — you now need to consider that your user may have overridden the underlying model at the OS level. Apple has indicated that developers can specify capability requirements and that app-level model selection can coexist with system-level defaults. But the edge cases are real: an application built around one model's behavioral idiosyncrasies should be tested against all three likely substitutes before shipping into the iOS 27 ecosystem.
The distribution question for AI labs just entered a new phase.
The short of it.
At WWDC 2026 on June 9, Apple announced AI Extensions — a framework in iOS 27 that lets users set Claude, Gemini, or ChatGPT as the model behind Siri and all Apple Intelligence features, routing queries system-wide without opening a separate app. Apple retains control of the interface, privacy handoff, and the settings page where users choose. Simultaneously, Apple is offering free access to its own models for small developers. For AI labs, this is the distribution channel they couldn't build; for app developers, it means the model your product was designed for may be overridden at the OS level — test your behavior across all three major providers before iOS 27 ships.