TL;DR / Key Takeaways
- Two weeks ago we mapped what iOS 27 would let developers build.
- Apple gave the WWDC 2026 keynote on June 8 — so we graded ourselves.
- Both big bets landed: the Foundation Models framework now takes image input (the 1–2GB bundled-VLM tax is over), and App Intents became the way into Siri (with SiriKit now deprecated on a 2–3 year clock).
- We even called the Core AI rename.
- What we missed: Apple outsourced Siri's brain to Gemini (~$1B/year) and made MCP system-wide across iOS 27 and macOS 27.
- The reality map for what to build now that the rumors are facts.
On May 26 we published a builder's opportunity map for WWDC 2026. The bet was simple: ignore the consumer polish, and watch two rumors that would actually change what an indie developer should build. We stated our confidence levels in plain text so you could grade us later.
Apple gave the keynote this morning — June 8, at Apple Park. So let's grade us. No grading on a curve: here's exactly what we said, exactly what shipped, and — more usefully — what the resolved reality means for what you build this week now that the rumors are facts.
The scorecard
Everything we flagged, scored against the keynote and the developer sessions:
| Our May call | Stated confidence | What Apple shipped June 8 | Grade |
|---|---|---|---|
| On-device Foundation Models go multimodal — image input for third-party apps | Medium — "the whole ballgame" | Shipped. The Foundation Models framework now accepts images alongside text; the new on-device model is natively multimodal | ✅ Hit |
| Core ML reframed toward a broader "Core AI" framework | Medium | Shipped, by that exact name — Core AI succeeds Core ML, with dynamic on-device/cloud/Extension routing | ✅ Hit (called the name) |
| App Intents gets a major Siri push; it's the highest-ROI thing to adopt now | High | Shipped — and harder: SiriKit was formally deprecated; App Intents is now the only path to the new Siri | ✅ Hit, then some |
| Siri "Extensions" — route Siri to third-party models like Claude/Gemini | Medium | Shipped — an Extensions picker in Settings: Gemini (default), Claude, ChatGPT, across Siri, Writing Tools, Image Playground, Xcode | ✅ Hit |
| The "noise" list (AirPods settings, casting, Genmoji, Siri visual redesign) won't change a roadmap | n/a | All shipped, all cosmetic for builders: AirPods custom EQ, EU Google Cast default, Image Playground bumps, Liquid Glass opacity slider | ✅ Correctly dismissed |
| Apple hasn't ruled out charging for deep Siri integration — don't bet your roadmap on the terms | Hedge | No developer charge announced; App Intents adoption is free | ✅ Hedge paid (resolved in your favor) |
| Watch for a device/chip floor on the new on-device AI | Watch item | Apple confirmed a "higher-power" model for some devices but didn't say which — still unresolved | 🟡 Flagged, unanswered |
| (Not predicted) MCP goes system-wide; Siri's brain becomes Gemini | — | Both happened — the two biggest structural moves of the day | ❌ Missed |
What we got right — and what it now means you can do
1. The cost wall fell. Delete the bundled VLM.
This was our headline bet, at medium confidence, and it landed. The Foundation Models framework — the third-party developer surface, the one you call from Swift — now accepts image input. You pass an image alongside text into the on-device model. Apple also threw in three things we didn't even ask for: custom skills, server-side model execution, and on-device fine-tuning.
The flip side we warned about is also live: if your edge was the hard VLM-bundling work, Apple just commoditized it. Your moat has to move up-stack — to the specialized reasoning, the data, the workflow — because the raw "look at this image on-device" capability is now free and built in. Our May advice — isolate your AI layer behind an adapter so the swap is a one-file change — is now the actual task in front of you, not a hypothetical.
2. "Core AI" — we called the name
We wrote that reporting pointed to "a reframing of Core ML toward a broader Core AI framework." Apple shipped exactly that, under exactly that name. Core AI succeeds Core ML, with native streaming token generation and a routing layer where your app describes a capability requirement — latency, privacy, capability level — and Core AI decides whether to serve it from the on-device model, Apple's cloud, or the user's chosen Extension. You stop hard-coding which model runs; you declare intent and let the OS route. That's a bigger idea than we predicted, but the name and the direction were right.
3. App Intents was the right thing to adopt — and now it's mandatory
We called App Intents "the single highest-ROI pre-WWDC move" at high confidence. Correct, and Apple raised the stakes: SiriKit is now formally deprecated, with a reported two-to-three-year window before SiriKit apps lose voice-assistant functionality entirely. App Intents is the only path into the new Gemini-powered Siri.
App Intents 2.0 is the surface we described — your app becomes something Siri calls directly, without the user opening it — now with richer entity types, streaming responses, and conversational follow-ups, so Siri can hold a multi-turn dialogue inside your app instead of dumping the user on a results card. The framing from the developer briefings is blunt: if Siri can't call your app, you're invisible in the agentic App Store. That's the same point we made in May, with a deprecation clock now attached to it.
4. The noise was noise
We listed the AirPods settings revamp, default non-AirPlay casting, Genmoji/Image Playground bumps, and the Siri visual redesign as real-but-irrelevant to a build decision. All shipped — AirPods got custom EQ and a reorganized menu, EU users can set Google Cast as the default casting protocol, Image Playground got a performance pass, and Liquid Glass got a transparency slider. Nice for users. None of it opens an app category. (Our one imprecision: we said "Maps over satellite"; what actually shipped was a Flyover upgrade. Same bucket — cosmetic for builders.)
What we missed — and these two matter most
Being right about the buildable stuff is the job. But the two largest moves of the day weren't in our map, and one of them is squarely in our lane. Honesty first:
Miss #1: Siri's brain is now Gemini
We framed Siri Extensions as a way for you to route Siri to an external model. We did not predict that Apple itself would route Siri to an external model — that the new, rebuilt, standalone-app Siri would run on Google Gemini, under a deal reported at roughly $1 billion per year. Apple's own foundation models still run on-device and on Private Cloud Compute, and Apple kept hammering privacy as non-negotiable, but the conversational brain of its flagship assistant is now Google's. That's the strategy story of the decade and we under-weighted it. For builders the practical read is unchanged — you reach Siri through App Intents either way — but it tells you how seriously Apple is treating the model layer as someone else's problem, and how much room that leaves above the model for third-party apps.
Miss #2: MCP went system-wide — and this one's ours to own
This is the miss that stings, because it's our home turf. Apple made the Model Context Protocol a platform feature. MCP support — which first showed up in Xcode 26.3 back in February — now extends across iOS 27 and macOS 27. MCP servers can be invoked by the new Siri and the Core AI routing layer; users (or enterprise MDM) register them explicitly, and from there your MCP server is a first-class thing the OS can call.
We didn't have MCP anywhere in the May post. If we rewrote that post today, "expose your product as an MCP server" would sit right next to "adopt App Intents" at the top of the list — because as of this morning they're the two supported ways to be callable by the agent on a billion devices.
The smaller misses, for completeness
- 1Agentic Xcode. Apple's IDE now runs an AI coding agent that can use the model and agent of your choice, handle localization, and drive simulated devices — Federighi called Xcode "the best place" to build apps. We said nothing about the build tools getting agentic. Worth a look this week.
- 2The speed story. iOS 27's marquee non-AI pitch was performance — ~30% faster app launches, 80% faster AirDrop, 70% faster photo loading — and the widest device support ever, back to the iPhone 11. We treated the keynote as an AI event; Apple also made it a "your old phone feels new" event.
- 3Tim Cook is leaving. He announced this was his last WWDC as CEO; John Ternus takes over September 1. Not a build decision, but the biggest headline of the day, and not on anyone's bingo card for the keynote stage.
The reality map: what to build now that the rumors are facts
In May the right column of our table said "build now" or "wait for June 8." June 8 happened. Here's the same map with the conditionals resolved:
| Opportunity | May verdict | Now that it shipped |
|---|---|---|
| Private on-device image understanding (scan a label, doc, receipt — never leaves the device) | Prototype on a bundled VLM; swap if the API ships | Build it on the system API. Image input is live in Foundation Models. Delete the bundled weight; ship a smaller binary. |
| Siri-callable actions in your app ("add this," "log that," "start this") | Build now — App Intents is shipping today | Mandatory, with a clock. Migrate off SiriKit to App Intents. Adopt App Intents 2.0 streaming + follow-ups. |
| Bring-your-own-model assistant (route a task to Claude/Gemini) | Wait for the API + terms | Buildable. Extensions shipped with Gemini, Claude, and ChatGPT selectable. No developer charge announced. |
| Offline/airplane-mode AI (summarize, classify, extract, no network) | Build now; a better model upgrades it free | Do it. The on-device model is bigger and multimodal; Core AI routes privacy-sensitive work on-device by declaration. |
| Expose your product to the OS agent (NEW — we missed this in May) | — | The new top-of-list move. Ship an MCP server and/or App Intents. iOS 27 + macOS 27 can now call both. |
| A tracker of which apps adopt App Intents / MCP / Extensions | Wait — no adoption data until ~Sept | Now buildable as the betas land. Adoption data starts accumulating this summer; stable ships ~September with iPhone 18. |
The deadlines that now exist
The single most useful thing the keynote did was convert soft "someday" pressure into dated deadlines. Put these on a calendar:
- 1SiriKit deprecation clock: ~2–3 years. Apps still leaning on SiriKit lose voice-assistant functionality when the window closes. Migration to App Intents is no longer optional; it's scheduled.
- 2Betas: today. iOS 27 / macOS Golden Gate developer betas are out now. The Foundation Models image API, Core AI, App Intents 2.0, and system MCP are testable this week, not in September.
- 3Stable: ~September 2026, with iPhone 18. That's when your users actually get it — and when adoption data becomes real. The team that shipped App Intents + MCP support during the beta is the one that's callable on day one.
- 4EU caveat. The new Siri is English-first and reportedly delayed on iOS/iPadOS in the EU under the Digital Markets Act. If your market is European, the App Intents/MCP plumbing still matters; the Siri front-end may lag.
The pattern held
We said in May that Apple's gaps are where indie opportunity lives, and that keynotes are where those gaps open and close. The on-device vision gap was the one to watch — and Apple closed it for you, which is the good outcome: the capability is now free, and the competition shifts to what you build on top of it.
The part we under-called is that Apple didn't just close a gap — it opened a new front door. Between App Intents and system-wide MCP, the iPhone is now an agent that can call your software directly. The apps that win the next year are the ones that are callable: by Siri, by Core AI, by any agent that speaks MCP. We'll be tracking exactly that — which products become callable, and which stay invisible — because that's the signal that now decides distribution. Grade us again next year.
Frequently asked questions
What did Apple actually announce at WWDC 2026 for developers?
The big four: (1) the Foundation Models framework gained image input, so third-party apps can run multimodal AI on-device without bundling their own vision model; (2) Core ML was renamed and reframed as Core AI, with a routing layer that sends work to on-device, cloud, or a user-chosen external model; (3) SiriKit was deprecated in favor of App Intents 2.0, now the only path into the new Siri; and (4) the Model Context Protocol (MCP) became a system-wide feature across iOS 27 and macOS 27. The consumer headline was a Gemini-powered Siri rebuilt as a standalone app.
Did Apple's on-device model get image input, like the rumors said?
Yes. As of iOS 27, the Foundation Models framework accepts image input alongside text, and the new on-device model is natively multimodal. This is the change that lets developers drop the 1–2GB vision-language models many apps bundled by hand and call a system API instead — smaller binaries, no first-launch download, and free model upgrades each OS release.
Is SiriKit really being deprecated?
Yes. Apple issued a formal SiriKit deprecation notice at WWDC 2026, with a reported two-to-three-year window before SiriKit-based apps lose voice-assistant functionality. App Intents is the replacement and the only framework that connects your app to the new Gemini-powered Siri. If your app uses SiriKit, migrating to App Intents is now a scheduled requirement, not an optional upgrade.
Can users choose a different AI model for Siri now?
Yes. Apple shipped an Extensions framework that lets users pick their AI provider in Settings — Gemini is the default (and powers Siri under the hood via a reported ~$1B/year Google deal), with Anthropic's Claude and OpenAI's ChatGPT also selectable. The choice applies across Siri, Writing Tools, Image Playground, and Xcode's contextual AI.
What is MCP and why does Apple supporting it matter?
The Model Context Protocol (MCP) is an open standard for letting AI agents call external tools and data sources. At WWDC 2026 Apple made MCP a system-wide feature in iOS 27 and macOS 27: registered MCP servers can be invoked by Siri and the Core AI routing layer. Practically, that turns the iPhone into an MCP client — if you run an MCP server, Apple devices can now call your tools directly, making it a real distribution channel alongside App Intents.
What should an indie developer build first after WWDC 2026?
Three moves, in order. First, migrate to App Intents (the SiriKit clock is running) and adopt the 2.0 streaming and follow-up features. Second, if you have private-vision features, rebuild them on the Foundation Models image API and delete any bundled vision model. Third — the one most teams will skip — expose your product as an MCP server so the OS-level agent can call it. The betas are out now; the goal is to be callable on day one when the stable release ships in September.