Context where the work happens

Picture someone pasting a paragraph into a chatbot so it knows what they mean. Apple Intelligence skips that step. It works from what’s already on screen, in system search, in the camera or in personal apps.

• Onscreen awareness
• Personal context across supported apps
• Visual Intelligence

A platform layer, not a universal workbench

For deep research, big deliverables and complex enterprise orchestration, a specialist work agent may still be the better place. Apple’s strength is integration and context. And server-side features now carry daily usage limits. A heavy production workflow can hit them.

• Use specialist agents when needed
• Keep the Apple layer as the entry point
• Design clean handoffs

Supported iPhone baseline

Apple lists iPhone 16 models or later, plus iPhone 15 Pro and Pro Max, for Apple Intelligence and Siri AI. The September 9, 2026 line adds iPhone 18 Pro and iPhone 18 Pro Max on the A20 Pro. It also adds the foldable iPhone Duo, on the same chip. Treat the performance figures as vendor claims, not independent measurements.
· Apple on the A20 Pro: “a new 6-core CPU, 7-core GPU, and Dual 16-core Neural Engine” with “double the AI processing power of A19 Pro.”
· Apple on iPhone Duo: it delivers “up to 35 percent better sustained performance than iPhone 17 Pro.”

• iPhone 18 Pro and Pro Max went on sale Friday, September 18 in 65 or more countries, with additional regions on September 25
• iPhone Duo is $1,999 for 256GB, pre-order October 16, available October 23, shipping on iOS 27.1
• Check exact model, storage, language and region before rollout

Three different Apple Watch floors

This one catches people. Apple Intelligence on Apple Watch needs Series 9 or later, or Ultra 2 or later, plus a nearby Apple Intelligence-enabled iPhone. Siri AI reaches Apple Watch Series 12 and Apple Watch Ultra 4 with watchOS 27. And Audio Intelligence sits higher again. The footnotes to the September 14, 2026 Siri AI release require Series 12 or Ultra 4. Live Rewind and Siri Recap sit on top of that, in beta in late 2026. Apple Vision Pro has its own spatial Siri AI experience. And it’s available in the EU.

• Series 9 or Ultra 2 for Apple Intelligence; Series 12 or Ultra 4 for Audio Intelligence
• Watch inherits paired-iPhone availability
• Watch is delayed in the EU; Vision Pro is not

Five metered features, and one of them is other people’s apps

Apple Intelligence isn’t clearly free any more. And the metered list is longer than the marketing page suggests. Apple’s dated support page of September 9, 2026 names five groups. Siri AI. Intelligent Photo Editing Tools, meaning Clean Up, Extend and Spatial Reframing. Image Playground. AFM 3 Cloud and AFM 3 Cloud Pro models in Shortcuts. And developer apps using Apple Foundation Models in Private Cloud Compute. Read that last one twice. A third-party app that calls PCC burns the same user allowance Siri AI does.
· Apple on what varies: the limits “may vary by feature, request complexity, system demand, system policies, and other factors.”
· Apple on recovery: a feature you have exhausted becomes available again after a time out period, which Apple does not quantify.
· Apple on cost, on the support page without the hedge the marketing page carries: “Increased access to such features will be available for a fee.”

• No request count, no token count, no reset window, no price
• A line-of-business app calling PCC shares the user’s allowance
• Budget a fallback path for a workflow that can exhaust an allowance

Apple’s two Rosetta notices disagree

Pick the wrong one and you plan the wrong year. Apple Developer News on September 1, 2026 says macOS 27 is the final release to support Rosetta. It adds a carve-out. Rosetta will keep working for older, unmaintained games that rely on Intel-based frameworks. The September 9, 2026 post says something else. It says macOS 26 was the final release for Intel Macs and Rosetta, and macOS 27 is Apple silicon only. Both posts are live. And macOS 27 shipped on September 14, 2026, so this isn’t academic. It decides whether your Intel-only app still runs on macOS 27. Test on a real macOS 27 machine. Trust the test, not either post.

• Two live Apple posts name different final-Rosetta releases
• Inventory Intel-only apps and test them on shipped macOS 27
• Both posts sit on a rolling index; confirm them directly

The September 9 hardware line

Apple titled the event “Surprise and shine.” It produced iPhone 18 Pro and iPhone 18 Pro Max, the foldable iPhone Duo, AirPods 5, Apple Watch Series 12 and Apple Watch Ultra 4. Apple published four dated newsroom releases that day. They cover the iPhone 18 Pro pair, iPhone Duo, AirPods 5 and Apple Watch Ultra 4. Apple Watch Series 12 got no release of its own. It shows up inside the Ultra 4 release and in the September 11 retail preview. So cite the dated press releases, not an event recap.

• The dated releases, not an event page, are the evidence
• Apple Watch Series 12 is cited inside the Ultra 4 release, not through one of its own
• Hardware performance figures in them are vendor claims
• The lineups reached stores on Friday, September 18, 2026 in 65 or more countries, with further regions on September 25

Observe, interpret, record

Use the camera to identify or understand something in front of you. Then create the note, task or message while the evidence is still there.

• Maintenance and inspection
• Retail and merchandising
• Site visits and logistics

Glance, question, delegate

Start from the alert or document already on screen. Ask what it means. Then start the next approved action or handoff.

• Decision preparation
• Mobile approvals
• Continuity into Mac workflows

The five models, named and sized

Apple built its name on doing things on the device. Its newest models were built with Google, and Apple documents them itself. The June 8, 2026 research post names a family of five, custom-built in collaboration with Google. AFM 3 Core is a 3 billion parameter dense model. AFM 3 Core Advanced is a 20 billion parameter sparse model. It activates 1 to 4 billion parameters per request. AFM 3 Cloud and ADM 3 Cloud (Image) run on the server, the second for image generation and editing. AFM 3 Cloud Pro is the most capable. It’s the one that moved onto Google’s hardware.
· Apple’s wording: “For AFM 3 Cloud Pro, we worked with Google and NVIDIA to extend Private Cloud Compute to NVIDIA GPUs in Google Cloud, while maintaining the same guarantees to protect our users’ privacy.”

• Five models, not one
• AFM 3 Core Advanced is what the 12GB unified-memory floor gates
• AFM 3 Cloud Pro is the model running outside Apple’s data centers

Private Cloud Compute is no longer only Apple’s hardware

This is the fact to carry into a design review. Apple says it’s expanding Private Cloud Compute beyond its own data centers. It’s working with Google and NVIDIA to run new Apple Intelligence workloads on Google Cloud. The silicon named isn’t Apple’s. It’s NVIDIA Confidential Computing with NVIDIA GPUs, Intel CPUs with TDX, and Google’s Titan chip. Apple says the five core PCC requirements still hold. Those are stateless computation, enforceable guarantees, no privileged runtime access, non-targetability and verifiable transparency.
· How Apple says it keeps the last one: it maintains “a cryptographically verifiable, append-only ledger of all Google Cloud hardware that is part of the PCC fleet.”
· How outsiders can check: Apple offers “access to live PCC nodes in research mode through the Apple Security Bounty Program.”

• A PCC request may execute on NVIDIA GPUs in Google Cloud
• The framework exposes PrivateCloudComputeLanguageModel and a dedicated PCC entitlement
• Ask for the hardware ledger and the research-mode evidence, not the summary

Research publishing is ahead of the API reference

Apple documented AFM 3 in June and hasn’t caught the framework docs up. Read on October 1, 2026 through the markdown endpoint, the Foundation Models updates page still tops out at a June 2026 section. The HTML version needs JavaScript and returns nothing. That June section does cover the 27-series. It tells developers to test prompts again, because the model changes when a person updates to iOS 27, iPadOS 27, macOS 27 and visionOS 27. What it doesn’t do is name AFM 3, its five members or their sizes. Those live on Apple’s research site. So the gap isn’t missing documentation. It’s two Apple publications moving at different speeds. You need both open.

• Model names and sizes: the June 8, 2026 research post
• API behavior and prompt-retest advice: the framework docs
• Don’t assume a documented API surface exists for every model Apple names

On-device first

Local processing can keep eligible data and inference on the device. That cuts data movement and keeps things quick. Two keys in Apple’s intelligence declaration make it a setting, not a hope: ForceOnDeviceOnlyDictation and ForceOnDeviceOnlyTranslation.

• Know which requests stay local
• Hardware and memory affect capability
• Device security remains essential

Google is inside the default model layer

Apple’s privacy reputation rests on keeping work on the device. Its newest foundation models are a family of five “custom-built in collaboration with Google.” And the most capable one runs on Google’s cloud. So the privacy review has to change. The outside-provider question isn’t confined to an opt-in routing path any more. It sits in the default layer that Siri AI and app intelligence use. Here’s what Apple does say. The models are Apple’s. The PCC guarantees hold. The Google Cloud hardware is logged in a verifiable ledger. Here’s what it doesn’t say. It doesn’t say what data, if any, Google as a company can see. Or whether any enterprise data-handling commitment differs from the consumer one.

• Record the collaboration and the Google Cloud hardware in the data-flow review
• Apple’s published promises describe the architecture, not a separate enterprise contract
• Ask Apple directly before signing a data-handling attestation that depends on it

There is now a real MDM control surface

Stop writing Apple Intelligence policy you can’t enforce. An admin can now push one declaration that blocks individual Apple Intelligence features, or keeps dictation and translation on the device. It’s com.apple.configuration.intelligence.settings. The table below sets out where it applies. The keys in Apple’s own example are AllowAppleIntelligenceReport, AllowGenmoji, AllowImagePlayground, AllowImageWand, AllowPersonalizedHandwritingResults, AllowVisualIntelligenceSummary, AllowWritingTools, ForceOnDeviceOnlyDictation and ForceOnDeviceOnlyTranslation. A nested Apps object adds per-app controls. They cover Mail smart replies and summaries, Notes transcription and transcription summaries, and Safari summaries.

• Supervised enrollment only, from the 26.4 releases
• Multiple configurations combine into one effective configuration
• Per-app keys let you allow summaries in Safari and block them in Mail

Two provenance mechanisms, pulling different ways

Apple is shipping provenance twice, for two different problems. A compliance program needs both on the page. SynthID marks images generated or edited with Apple Intelligence tools. Apple dates it to later this year. Apple Reference Image goes the other way. It’s an opt-in camera mode that makes securely timestamped reference images, using on-device secure hardware and Private Cloud Compute. It debuts on the main camera sensor of iPhone 18 Pro and iPhone 18 Pro Max. Apple published it on September 15, 2026, with a critique of the incumbent standard. Apple argues that C2PA-style approaches attach provenance metadata after capture. They certify the edit history only from that point on. So they’re open to compromise anywhere in the editing chain.

• SynthID marks AI-edited images; Apple Reference Image attests authentic capture
• Apple Reference Image is opt-in, iPhone 18 Pro and Pro Max only, and uses PCC
• If your program standardized on C2PA, read Apple’s objection before assuming alignment

Personal context boundary

The same device may hold personal and corporate information. And easy retrieval can mix the two in ways nobody intended.

• Use managed accounts and devices
• Set data-handling policy
• Avoid exposing unrelated context

Regional fragmentation

Apple says Siri AI isn’t available for iOS, iPadOS or watchOS in the EU. It is available on macOS 27 Golden Gate and visionOS 27. But the Apple Account region has to be in the EU, and the device and Siri languages both have to be English. Apple also says Siri AI and the other new Apple Intelligence features won’t be available in China while it works through regulatory requirements. So one global workflow can’t assume one feature set. Or even one language.

• Design capability detection
• Provide a non-AI path
• Review rollout by platform, region and language

Third-party model routing

A common developer framework doesn’t make all models equal on privacy, retention, security or quality. That was already true of the LanguageModel protocol route. Now it’s true of the default route too. Apple describes its own foundation models as built with Google, and it runs the largest one on Google Cloud. So provider evaluation can’t stop at the edge of the Apple stack.

• Document provider selection for every route, including Apple’s own
• Evaluate each route independently
• Make external transfer and external execution visible

Beta expectations

Siri AI shipped on September 14, 2026, as a beta in English. Released isn’t the same as finished. A beta isn’t a production service level. Pilot it, measure it, and keep a path that works when it doesn’t.

A dated promise is still a promise

Apple named features that didn’t ship, and most now carry a date. SynthID support is dated later this year. Live Rewind and Siri Recap on Apple Watch are dated beta in late 2026. They won’t be available in the EU at first. The Health app redesign with an Insights tab has no date. Nor do suggestions in Messages conversations. And the Siri AI language expansion was dated October 2026 and hasn’t arrived.

Metered server capacity

Apple meters Siri AI, Intelligent Photo Editing Tools, Image Playground, and AFM 3 Cloud and AFM 3 Cloud Pro models in Shortcuts. It also meters developer apps calling Apple Foundation Models in Private Cloud Compute. It publishes no number for any of them. So a workflow that assumes unlimited Siri AI, Image Playground or in-app PCC calls has an untested failure mode. And your own app can be the thing that runs it dry.

No Apple page names a third-party assistant

The external intelligence declaration exists, with AllowSignIn and AllowedWorkspaceIDs. So Apple clearly expects third-party assistants. But no Apple page names one. The Apple Intelligence page mentions no third-party model at all. The September 14 newsroom releases name only Google and Gemini. If a rollout depends on a specific named assistant staying integrated in the 27-series, get that from Apple. Not from a press report.

Regional gaps

iPhone, iPad and Watch are excluded in the EU. Mac and Vision Pro aren’t, but there the EU path needs English for both the device and Siri. China is confirmed unavailable while Apple works through regulatory requirements. A workflow that looked universal in a headquarters demo can break at either border.

Version drift inside the launch

The 27.0 releases date from September 14, 2026. A 27.0.1 round shipped on September 28, and the 27.2 line is on beta 2. iOS 27.1 has no beta and no release. Yet it’s the build iPhone Duo ships on, October 23. So pin the build number in any test result or rollout plan. “iOS 27” now covers several behaviors.

Deprecated management profiles

Apple has deprecated the old restrictions for managing intelligent systems on-device. They cover Apple Intelligence, External Intelligence, Siri and Keyboard. So a fleet still enforcing Apple Intelligence through them is on a path Apple has marked for retirement.

Visual misinterpretation

Camera and screen understanding can’t stand in for required measurements, expert inspection or source documents.

Model change

OS updates can change on-device model behavior. Apple’s own June 2026 developer note says to retest prompts, because the model changes when a person moves to the 27-series. Now there’s final software to retest against. And it’s already at 27.0.1.