Tim Cook walked out on stage today for the last time as CEO of Apple. His farewell WWDC. His mic-drop moment. His legacy-cementation event. The press is already writing the tributes.
And what did Apple deliver to mark the occasion?
They paid Google — Google — roughly a billion dollars a year to make Siri not embarrassing anymore. They slapped a transparency slider on the interface disaster they shipped last year and called it a design improvement. And they brought out a medical professional to explain to a room full of developers that children, apparently, should not have unlimited access to TikTok.
This is what we waited two years for.
The Siri Situation: A $1 Billion Admission of Failure
Let's be honest about what happened here. Apple promised a context-aware, personally intelligent Siri at WWDC 2024. They plastered it on billboards. They used it to sell iPhone 16 units. Then they failed to ship it for nearly two years, settled a class action lawsuit for $250 million in May, and showed up today with a Google-powered replacement wearing an Apple badge.
The new assistant is called "Siri AI" — because evidently the branding team was also on vacation — and it runs on a custom 1.2-trillion-parameter Gemini model that routes heavy queries through Nvidia B200 GPUs on Google Cloud. This is the company that spent years telling you privacy is a product. This is the company that built Private Cloud Compute specifically so they'd never have to do this. And now your voice queries are going to Google's infrastructure because Apple couldn't build a competitive model on their own timeline.
To be fair to the product itself: it actually looks good. Multi-step chained commands. On-screen awareness. Personal context that can read your emails, photos, and files to answer questions. A standalone Siri app with conversation history. Deeper integration into Spotlight on macOS. Cross-device continuity so you can start a Siri conversation on your iPhone and pick it up on your Mac.
The infrastructure story is more nuanced than "your data goes to Google" — and it's worth getting right, because Apple actually built something architecturally interesting here. It's a three-tier routing system. Simple tasks run entirely on-device using Apple's own small models. Moderately complex requests go to Apple's Private Cloud Compute: Apple Silicon servers, stateless ephemeral processing, no data retained after a query resolves, independently audited by an ACM paper this month. Only the heaviest reasoning tasks — the ones that require the full 1.2-trillion-parameter Gemini model — route to Google Cloud, running on Nvidia Blackwell B200 GPUs with hardware-level confidential computing enabled. Queries are anonymized and stripped of Apple ID linkage before they ever get there. Apple's contract prevents Google from using any of it to train future models.
Here's the thing though: Apple tried to run Gemini entirely within Private Cloud Compute and couldn't. The model is roughly eight times larger than anything Apple had previously built, and PCC hardware wasn't fast enough to handle it at Siri's query volumes. So they bolted Google Cloud onto the side of their own privacy architecture because their own privacy architecture hit a wall.
The privacy properties are real — the independent audits are real, the contractual protections are real. But Apple spent years telling you Private Cloud Compute was the answer to cloud AI privacy, and then quietly had to admit it can't actually run the model they need at scale. They'll retain the PCC branding anyway, naturally.
This is the AI assistant Apple promised in 2024. It just took a failed internal program, an executive shuffle that pushed out John Giannandrea, and a last-minute $1B/year deal with the company Apple's supposed to be competing with to actually ship it. John Ternus, the incoming CEO, inherits whatever this arrangement actually becomes in practice.
The capability is real. The path to get here is a disaster they're hoping everyone forgets. They won't.
Liquid Glass: Now With a Slider You Can Use to Make It Less Terrible
Last year Apple introduced Liquid Glass, a unified design language inspired by visionOS that turned every UI element on your phone into a translucent smear. The reaction was, charitably, mixed. Less charitably, people found it actively difficult to read, particularly on anything that wasn't a brand new display in perfect lighting conditions.
Apple's solution? They added a slider.
You can now adjust how transparent Liquid Glass windows are. You can "diffuse complex content in the background to make text more legible." App icons got updated contrast. The design team apparently processed "a lot of user feedback" — which is corporate speak for "people were furious and we heard them."
This is fine. This is the right call. A new design language that ships rough and gets iterated on is better than a design language that ships rough and Apple doubles down on it out of pride. The original rollout was a mess, the complaints were legitimate, and the response — while overdue — is proportionate.
What I find funny is that "here's a slider to make our redesign less aggressive" is being framed as a feature announcement. This is an apology with a UI. And I'll take it. Accessibility and readability matter more than aesthetic purity.
macOS gets the same treatment, with Liquid Glass tweaks landing in macOS Golden Gate — a fine name, Craig delivered it with appropriate gravitas.
Think About the Children (No, Really, They Spent a Lot of Time on This)
Apple dedicated a noticeable chunk of the keynote to child safety features. A medical professional was brought out. Research was cited. The phrase "digital journeys" was used.
The actual features are reasonable. Parents can now set time limits by app category — Entertainment, Games, Social Media — instead of micromanaging individual apps. There's a Screen Time Schedule that controls which categories are available at specific times of day. Kids can "Ask to Browse" a website, which sends a permission request to parents before access is granted. Communication controls let parents define exactly who their children can talk to, with kids requesting additions to the approved contact list.
These are fine features. Genuinely useful for parents who are trying to navigate the impossible task of letting kids have devices while not letting devices have kids.
But the production value on this segment felt like regulatory pre-emption as theater. Apple is watching every Western government with a child-safety bill moving through parliament, and they're building a paper trail. "We didn't need a law to do this. Look, we brought a doctor." The features are real. The motivation is partly legal prophylaxis. Both things are true simultaneously, and I don't think acknowledging that makes Apple evil — it makes them a company doing the right thing for a mix of right and expedient reasons, which is basically how most good corporate behavior works.
What I'll actually watch for: whether any of this interoperates with non-Apple apps in practice. The communication controls framework requires third-party developers to adopt it. Instagram isn't going to do that. The gap between "Apple built the scaffolding" and "children are actually safer" depends entirely on an ecosystem compliance problem Apple can demo but not mandate.
The Other Use Case Nobody Mentioned
Every feature in that child safety segment has a second application Apple didn't discuss.
Approved contacts lists — controlling exactly who someone can communicate with — is a textbook coercive control pattern. "Ask to Browse" requiring permission before visiting a website is how you monitor and restrict someone's access to information. Screen Time schedules that gate which apps are available at which hours of the day. Location visibility. Communication logging. All of it baked into the OS, all framed as parental tools, all functionally identical to the infrastructure an abusive partner, a controlling parent of an adult child, or an employer with a managed device would want.
Family Sharing has been used in domestic abuse contexts for years. Location sharing is the most documented case. The communication controls announced today are the same architecture. And because they're OS-level features rather than third-party apps, they carry Apple's implicit legitimacy instead of the red-flag visibility of stalkerware. They're harder to identify. Harder to remove. A victim who doesn't know what to look for won't find it in the App Store — it's just a Settings menu.
Apple brought a doctor on stage to frame these features as scientifically grounded child development tools. None of the research cited accounts for households where the person holding the Family Sharing account is the threat. The "child" framing does a lot of work to make the whole segment feel unambiguously good. It shouldn't go unexamined.
The Quietly Biggest Announcement Nobody Is Talking About
While everyone was watching the Siri demo and doing takes about Google, Apple slipped in the thing that will actually matter most to developers: the on-device Foundation Models are now multimodal.
The previous on-device model was text-only. Apple Foundation Models 2 — co-developed with Google — can understand speech, read text, and process images, all running locally on the device. That's not a footnote. That's an architectural shift.
Here's why it matters. Right now, every iOS app that does anything with visual AI faces the same ugly tradeoff: bundle a small, heavily quantized on-device model and accept capability limits, or make cloud calls and accept latency, cost, and a dependency on connectivity. Neither is great. The bundled-model path keeps apps lean but caps what the model can actually do. The cloud path works until it doesn't — no signal, API costs, privacy questions, round-trip latency on every inference.
A multimodal Foundation Models framework collapses that tradeoff. A system API for image understanding that any app can call, running on Apple's own optimized stack on-device, means developers don't have to choose between capability and bundle size. They get a capable vision model for free, at system speed, without a network call. Apps that were phoning home for visual inference can stop. Apps that were shipping stripped-down models because they couldn't afford the binary size for something better get an upgrade they didn't have to build.
The accessibility implications of this are real and underappreciated. AT apps that do visual interpretation — whether that's identifying objects, reading screenshots, describing UI elements — are currently either shipping large bundled models or making network calls. Neither is great. An on-device multimodal system API means that capability is available offline, at system speed, to any app that asks for it, with no additional binary size cost. That's the kind of infrastructure improvement that makes an entire category of assistive tools better without requiring each developer to independently solve the same problem.
Apple didn't headline this. They buried it in the architecture slide behind the Siri demo. But if you're building anything that needs to see the world — and in accessibility tech, a lot of things do — this is the announcement that changes your roadmap.
Spatial Reframing: Apple Finally Touches AI Photo Editing, Immediately Does Something Weird and Interesting
The consumer-facing proof of concept for the new multimodal stack is a feature called Spatial Reframing, and it's the kind of thing that sounds like marketing nonsense until you understand what's actually happening under the hood.
The pitch: you took a photo. The composition is wrong. Someone's head is cut off, or you were standing too close, or you wish you'd stepped left two feet. Spatial Reframing lets you adjust the angle or composition after the fact. Not a crop. Not a zoom. A generated version of the photo as if your camera had been positioned differently when you took it.
Apple has been notably, almost aggressively reluctant to apply generative AI to photos. While Google was shipping Magic Eraser and Move and Best Take and every other computational photography party trick, Apple kept the camera clean. The implicit pitch was that iPhone photos are real. They're not composited. They're not generated. You pointed a lens at a thing and here is what the lens saw.
Spatial Reframing is Apple blinking on that. They're generating pixels that didn't exist in the original capture. The presenter called it "like I was able to go back in time and adjust my camera in the moment," which is either a charming description or a slightly alarming one depending on how you feel about the indexical relationship between a photograph and reality.
What makes it interesting technically is that it's not just inpainting the edges. Apple is using 3D modeling — reconstructing a spatial understanding of the scene from the 2D image — and then rendering a new viewpoint from that model. That's a harder problem than most computational photography tricks. It's also a direct application of the multimodal Foundation Models that can now understand images at a structural level, not just classify them.
Whether the output is actually good in practice remains to be seen. Generating plausible geometry for regions the original camera never captured is exactly the kind of thing that looks impressive in a controlled keynote demo and falls apart on a photo of your cluttered kitchen. But the ambition is real, and it's a more honest use of generative AI in photography than pretending the feature is just "enhancing" what was already there.
The Developer Tax
Xcode 27 has an agentic coding assistant that can simulate entire apps. Federighi declared Xcode "the best place for agentic coding" to a room full of people who've been using Cursor or Claude Code for months. Write a prompt, watch it build and run in the simulator, iterate without touching the keyboard. Code completion runs on-device.
The Foundation Models framework now accepts image input for third-party developers, with custom skills and server models available too. Same on-device capability powering Spatial Reframing and the new Siri, callable by any app in three lines of Swift. No API key. No network call.
SiriKit is deprecated. App Intents is the only integration path going forward, with a two to three year migration window. App Intents lets Siri call into your app's actions and fetch real data. The new Siri finally has the reasoning capacity to use those integrations. Apple's deprecation timelines have a long history of outlasting developer compliance.
macOS Golden Gate drops Intel support entirely. Apple Silicon is three years mature. Intel Macs are a rounding error. The call is correct.
The Rest
Apps open 30% faster. Performance improvements are unglamorous and extremely worth doing.
The Passwords app now uses Apple Intelligence to agentically navigate websites and change insecure passwords on your behalf. Specific task, clear scope, concrete outcome. This is what useful agents actually look like. More of this.
Visual Intelligence is moving out of the Camera Control button — where apparently nobody found it — and into the Camera app as a proper Siri mode alongside Photo, Video, Portrait, and Panorama. Discoverability is a feature.
How Accessible Is the New Siri, Actually
Apple announced real accessibility improvements ahead of this keynote. VoiceOver gets richer image descriptions. Magnifier gets voice-controlled zoom and natural language queries about what the camera sees. Voice Control moves from rigid command syntax to natural language — "tap the orange folder" instead of memorizing a label grid. These are genuine upgrades and they got a Global Accessibility Awareness Day press release three weeks before WWDC.
None of that answers the question I actually care about.
The new Siri is a chat interface. A text input field. A scrolling conversation history. A paperclip to attach files. A Dynamic Island prompt that expands when invoked. Apple spent a significant portion of today's keynote demoing this UI, and not once did anyone show it working with VoiceOver. Not once did anyone discuss whether a blind user can navigate conversation history, whether the Dynamic Island interaction is reachable, whether the chat input field behaves correctly with a screen reader, whether the attachment flow is accessible, whether any of the visual affordances that make the interface intuitive for sighted users have accessible equivalents.
This matters because the old Siri interaction model was voice-first by design. You invoke it, you speak, it responds. The new model is chat-forward. It has been designed around a visual paradigm — the same paradigm as ChatGPT and Gemini — and then presumably audited for accessibility afterward, if at all. That sequencing is exactly how you get screen reader support that technically passes VoiceOver's focus order checks while being practically unusable by anyone who actually relies on it.
Apple's accessibility press releases are about VoiceOver describing the world more richly. That's AT built on top of the OS. The Siri app is the OS. Whether the company that built the best screen reader in the consumer market actually built its flagship new AI interface to work with that screen reader got zero stage time today.
The developer beta dropped today. The public beta follows in July. This is exactly what that process is for — blind users and AT professionals getting hands on the software and filing the radars before September. Apple's beta cycle has caught real accessibility regressions before they shipped. The question is whether enough of the right people are in the beta, whether the feedback channels are responsive, and whether there's time to fix what's broken.
Siri AI is itself labeled a beta within iOS 27 and ships behind a waitlist. So being in the iOS 27 beta doesn't get you access to the thing that actually needs testing. I'm in the beta. I've joined the waitlist. We'll see.
Who Gets What
iOS 27: iPhone 11 and newer — same support window as iOS 26, no cuts this year. Apple Intelligence and the new Siri require iPhone 15 Pro or newer (A17 Pro chip, 8GB RAM minimum).
macOS Golden Gate: Apple Silicon only. Intel is done. Four models that ran Tahoe don't make the cut.
watchOS 27: Series 10 and newer, Apple Watch Ultra 2 and newer. Drops the SE (2nd gen), Series 6, 7, 8, 9, and the original Ultra in one go — one of the biggest watchOS compatibility cuts ever.
iPadOS 27: Drops the iPad Air 3, iPad mini 5, and the 2018 iPad Pro (A12X). Supported from the iPad Pro A12Z (2020) and iPad Air 4 onwards. Apple Intelligence requires M1 or A17 Pro with 8GB RAM, with the new features requiring M4. Some are saying that M1 iPads are also not supported, but Apple's website disagrees.
Siri AI: English only at launch. And here's the one that deserves its own paragraph.
Apple spent today's keynote telling you Siri AI was built privacy-first, on-device where possible, anonymized at every tier, independently audited, contractually protected from Google training on your data. The whole thing. Privacy is the product.
Siri AI will not be available in the EU on iOS 27 or iPadOS 27 at launch. Apple is blaming the Digital Markets Act. The regulations designed to protect user privacy and give users more control over their data are, apparently, the reason Apple can't ship its privacy-first AI assistant there.
Apple's actual stated reason is more specific: under the EU's interpretation of the DMA, Apple would have to give any virtual assistant direct access to users' private data and the ability to directly control other installed applications the moment Siri AI becomes available. Which would mean dismantling the privacy architecture they just spent an hour explaining to you.
So the situation is this: Apple built a privacy architecture specifically to avoid handing user data to third parties. The EU built regulations specifically to force interoperability between assistants. Those two things are, at least according to Apple, incompatible. You can disagree with Apple's read of the regulation. But the tension is real, and it doesn't resolve cleanly into a punchline.
It will be available in the EU on macOS, visionOS, and watchOS. The two platforms most people actually use, iOS and iPadOS: not yet. No timeline.
China also doesn't get it, for separate regulatory reasons Apple didn't elaborate on.
The Verdict
This was a defensive WWDC. Apple spent two years hemorrhaging credibility on AI, settled a lawsuit about features they advertised and didn't ship, and showed up today with a rescue package largely built on Google's infrastructure. They patched their most embarrassing design rollout with customization options and parental controls they probably should have shipped a year earlier.
And yet. The features are real. The Siri overhaul is the most substantive upgrade to the assistant since it launched. The child safety work, however cynically motivated in part, is practical and thoughtful. The Liquid Glass retreat is the right call.
Tim Cook ends his WWDC run having presided over a two-year AI face-plant and its corporate-partnership-fueled recovery. John Ternus inherits a company that is, as of today, back in the game — or at least back at the table.
Whether Gemini-powered Siri ships on time, does what the demo showed, and actually makes people want to upgrade their phones: that's the only question that matters now. Apple has burned that particular trust once already this cycle. They don't get a third chance to promise something and not ship it.
We'll see in September.