Apple Intelligence adds Gemini-backed Siri beta with visual and on-screen understanding

• Apple says the next Apple Intelligence stack is built on Apple Foundation Models, while multiple WWDC posts, including testingcatalog's first report and testingcatalog's model stack post, say Gemini is part of the underlying model story.
• Apple positioned Siri AI as a later-this-year beta with personal context, broad web knowledge, onscreen awareness, and app actions, which lines up with Apple's Siri AI announcement and live posts from testingcatalog's Siri beta summary and testingcatalog's Siri app post.
• The most concrete new Siri behaviors in the evidence pool are visual understanding, camera mode, interactive widgets inside conversations, and promptable app actions, according to testingcatalog's onscreen understanding post, testingcatalog's widgets post, and testingcatalog's camera mode post.
• Apple separately confirmed that Siri AI will miss the EU launch for iOS 27 and iPadOS 27 in a DMA delay statement, echoing what testingcatalog and kimmonismus flagged during the keynote.
• Xcode 27 is part of the same AI reshuffle: testingcatalog's Xcode post spotted Claude and Codex in Apple tooling, and the Xcode 27 beta release notes add that Google Gemini is now available in the coding assistant too.

You can read Apple's main WWDC roundup, the dedicated Siri AI announcement, and the broader Apple Intelligence post. The weirder reveal is that the model story now sprawls from Apple Foundation Models to Gemini-backed infrastructure in keynote reporting, while Apple's own developer docs also put Gemini, Claude, and Codex inside Xcode on day one through the release notes. Apple also published a separate EU delay notice instead of burying that caveat in the launch post.

Apple's official framing is that the new system integrates its latest Apple Foundation Models deeply across the platform stack in the Apple Intelligence announcement. The live-keynote reporting was blunter: testingcatalog's first report and testingcatalog's model stack post both said the new Apple Intelligence is built on top of Apple Foundation and Gemini models.

That matters because Apple is no longer presenting Siri as a single-assistant upgrade. The company is describing a broader architecture change, and the external read on that architecture is now explicitly multi-model.

Apple's Siri AI launch post says the assistant gets personal context understanding, broad world knowledge, onscreen awareness, a dedicated app, and expanded Visual Intelligence. The keynote posts add the shipping shape: testingcatalog's Siri beta summary says the beta lands later this year, while testingcatalog's Siri app post and testingcatalog's conversational Siri post show Apple making the assistant look and sound more like an always-on chat surface.

The feature list in the evidence breaks cleanly into four buckets:

• Personal context in apps, according to kimmonismus.
• Web browsing and broad knowledge, according to testingcatalog's Siri beta summary and Apple's Siri AI announcement.
• Onscreen awareness, according to testingcatalog's onscreen understanding post.
• In-app usage and actions, according to kimmonismus and the WWDC26 App Intents session, which says developers expose content and actions to Siri through App Intents.

The most concrete demo thread in the evidence is Siri seeing what is on screen and what the camera sees. testingcatalog's onscreen understanding post shows the assistant answering a location question from visible content, while testingcatalog's camera mode post says Siri mode extends into Camera.

Apple's own wording in the Siri AI post matches that direction with onscreen awareness and expanded Visual Intelligence. The result is a Siri story that is less about voice commands and more about multimodal state.

The UI changes are unusually specific for an Apple assistant launch. testingcatalog's widgets post says Siri conversations can now contain interactive widgets and images, testingcatalog's Siri voices post says voices are customizable, and testingcatalog's conversational Siri post says Siri is much more conversational.

Apple's official post also says Siri AI gets a dedicated app for revisiting conversations across products. Put together, the assistant is being recast as a persistent interface, not a transient voice overlay.

Several of the most useful upgrades sit outside the assistant shell. testingcatalog's Safari features post lists three Safari hooks:

• Tab grouping
• A notify me feature
• Describe an extension

testingcatalog's Shortcuts post adds prompt-built shortcuts, and testingcatalog's thread context also points to suggestions in Mail and Calendar. Apple's Apple Intelligence post separately calls out smarter browsing in Safari, upgrades in Photos, and a rebuilt Image Playground.

Apple says new Siri AI features are available for developer testing now and will roll out later this year with the software releases, per the WWDC roundup and the Siri AI announcement. The bigger caveat is regional: Apple's separate DMA delay statement says Siri AI will not ship in the European Union with iOS 27 and iPadOS 27.

The evidence pool also points to a hardware split. kimmonismus said the new on-device Siri model is only on iPhone 17 Pro, then clarified in-thread that the restriction referred to the on-device model specifically. kimmonismus's EU reaction underlined the double caveat: high-end device requirement, then no EU launch anyway.

The last surprise is that Apple's AI reset is not confined to consumer surfaces. testingcatalog's Xcode post spotted Claude and Codex inside Xcode, and the Xcode 27 beta release notes say Google Gemini is also available in the coding assistant.

The same release notes mention Apple-authored agent skills and a workaround for exporting them to Codex, which is a more opinionated integration than a generic model picker. That makes WWDC's Gemini reveal bigger than Siri alone: Apple's own developer tooling now exposes multiple frontier assistants inside Xcode on day one.