Apple Siri overhaul and iOS 27 AI features at WWDC

The headline product story is the chat-style Siri app and Dynamic Island animation, but the structural story is that Apple has stopped pretending its in-house foundation models can carry the assistant. After the LLM-powered Siri promised at WWDC 2024 slipped repeatedly, Apple is reportedly paying Google roughly $1 billion per year to license a custom 1.2 trillion parameter Gemini variant that handles Siri's summarizer and planner functions ^[1]. The size of that licensed model — a trillion-parameter teacher — gives a sense of the capability gap Apple is trying to close from outside.

The consolation prize is architectural rather than narrative: Google's model reportedly runs on Apple's own Private Cloud Compute servers, so no user data is shared with Google ^[2]. Apple is also reportedly distilling the large Gemini model into a smaller version designed to run on-device on Apple silicon ^[3]. The strategic implication is that Apple's WWDC 2026 pitch is no longer 'we built the best model' — it is 'we built the best private surface for someone else's model, and we own the chip it runs on.' That is a more honest position than the 2024 keynote, but it also means Apple's AI roadmap is now tied to Google's release cadence in a way it has never been before.

Functionally, the new Siri app is an iMessage-style chat with persistent history, inline cards for weather, notes, and appointments, and a 'Search or Ask' panel that replaces Spotlight as the top-of-screen entry point ^[4][5]. Voice and text are equal-class inputs, attachments are supported, and an Extensions framework exposes a drop-down where users can route queries to ChatGPT, Claude, or Gemini instead of Apple's default ^[4]. None of those decisions are novel — they are the table stakes of a 2024-era chatbot app.

That is the point. For three years Apple positioned Siri as an ambient orchestration layer that did not need a chat surface. The redesign concedes that users actually want the chat surface, complete with rememberable threads and a drop-down model picker. The bet is that owning the OS-level entry point — Dynamic Island invocation, swipe-down 'Search or Ask', Camera-app Siri mode replacing the Visual Intelligence button ^[6]— turns competing assistants from destinations into plugins. Whether that distribution leverage is enough to claw back users who have already built habits inside ChatGPT and Claude is the unanswered commercial question of the keynote.

Apple is reportedly preparing to spend a meaningful share of WWDC 2026 stage time on on-device AI, leaning on 15 years of custom silicon as the differentiator ^[7]. Read against the rest of the industry's spend, this is less a technology claim than a financial one. The reporting frames on-device inference as a privacy and cost alternative to the massive data center buildouts rivals are pursuing, with cloud reserved for complex queries via Private Cloud Compute ^[3][7].

The distilled-Gemini-on-device approach is the technical expression of that bet, and it is a genuinely awkward one. A trillion-parameter teacher model distilled into something that fits within an iPhone's RAM and thermal envelope will lose substantial capability, and community technical commentary has been quick to note that even current Pro iPhones ship with 12GB of RAM, which is a hard ceiling on what local inference can usefully do. Apple's counterargument — that vertical integration of silicon, OS, and model lets it extract more useful work per watt than a generic stack — is plausible but unproven at the model sizes users now expect. The credibility of the WWDC pitch will rest on whether the on-device demos feel like real assistants or like throttled previews of the cloud version.

The reception around the May 28 leaks has been split along a predictable seam. The visual redesign — Dynamic Island animation, standalone app, dark color scheme matching Apple's WWDC art — is being received as a long-overdue UI refresh. The architectural disclosure that Gemini powers the underlying model is being received as an admission that two years of 'Apple Intelligence' marketing did not produce a competitive foundation model. Technical commentary in developer-leaning communities has been notably exhausted, with multiple threads pointing out this is reportedly Apple's fourth Siri redesign in roughly as many years.

The more interesting community thread is the one about what Apple should have done instead: bake a 15-30 billion parameter purpose-built model directly into an ASIC, trading generality for orders-of-magnitude efficiency on a narrow set of assistant tasks. That argument is a useful frame for evaluating the keynote. If Apple's on-stage story is 'we shrank Gemini onto your phone,' it concedes the architectural high ground to whoever ships the best generalist model. If the story is 'we co-designed silicon and a smaller model for the specific shape of assistant work,' it is a more defensible long-term position — and one no rival can copy without owning the chip.