Omni, Spark, a fall date for smart glasses–the keynote was a feature dump. The real story is the two things Google quietly took away.
After enough of these keynotes, you stop watching the slides and start watching the seams. A feature dump is just a feature dump. But when every announcement leans the same way–when Omni, Spark, Search, and a footnote about billing all point at the same horizon, the news isn’t on any slide. It’s the horizon. Google I/O 2026 was that kind of keynote, and the horizon it pointed at should change how you think about the phone already in your pocket.
The product list is real and worth having. Gemini Omni turns a selfie into a cinematic short and edits video through plain conversation. Gemini 3.5 Flash is fast, cheap, and good enough that Google is making it the default almost everywhere. Gemini Spark is a 24/7 agent. Audio glasses from Gentle Monster and Warby Parker arrive this fall. If you want that rundown in detail, it is everywhere by now, and most of it is accurate.
But the rundown is the surface. Underneath it, two things happened at I/O 2026 that Google never stated outright, because stating them plainly would have made the room uncomfortable. The first: the app stopped being the place where you get things done. The second: the assumption that a video shows something that actually happened stopped being safe. Both are consumer stories. Both change what the hardware you already own is for. Neither requires you to buy anything.
Here is a habit worth borrowing from years of doing this: ignore what an announcement is called and watch where it actually points.
Gemini Spark is called a personal agent. What it actually is, structurally, is a process that does not run on your phone. It runs on dedicated virtual machines in Google’s cloud, and Google says plainly that it keeps working “even when you close your laptop or lock your phone.” That sounds like a convenience line. It is actually a relocation. The thing doing the work is no longer in your hand. Your device became a window onto it.
Search made the same move from a different direction. Google’s Search team described Search building “mini apps” — trackers, dashboards, small tools — coded on the fly, per task, the moment you ask. A fitness tracker that pulls live weather and local reviews. A move-planning dashboard you return to for weeks. None of these live in an app store. None were installed. They are assembled when you need them and gone when you don’t. After more than a decade of watching the app-store model define how software reaches people, it was genuinely strange to watch Google casually invert it on stage.
And then the detail I think is the real headline, sitting in the subscriptions post where almost nobody looks: the Gemini app is moving from daily prompt limits to a compute-used model. You are no longer billed for how many times you ask. You are billed for how much work the machine does for you. A text question is cheap. A video edit or a coding task costs more. Your allowance refreshes every few hours against a weekly ceiling.
I have watched a lot of pricing-page changes get waved through as housekeeping. This one isn’t housekeeping. When a company stops metering access to a tool and starts metering outcomes, it has changed what it is selling. That is the same shift Spark embodies and the same shift the mini apps embody. For two decades the verb was open–open the app, open the doc, open the tab. Google spent I/O 2026 swapping it for ask, and quietly arranging things so that whatever happens after you ask happens somewhere you cannot watch.
It reaches the glasses too. The audio eyewear Google previewed will, in its own demo, line up your DoorDash coffee order before you have touched a phone–leaving you only the confirmation tap. Android Halo, the new status sliver at the top of the screen, exists for exactly one reason: when the app is no longer a place, you need a new way to even know a task is running. Halo is that gap, turned into UI. It is the most honest thing in the whole keynote.
For most people, day to day, this will feel like relief–less menu-tapping, more approving finished work. I am not here to talk anyone out of that. But I have learned to name the trade. An app you open is one you can see, inspect, and close. An agent on a cloud VM, billed by compute, working while your screen is dark, is a different relationship with your own hardware. Your gadgets did not get weaker this week. They got more dependent–on a connection, a subscription, and a company’s judgment about what to do while you are not looking.
The second removal is the one that, after years of writing about this stuff, genuinely gave me pause.
The same morning Google introduced Omni — a model built to generate and seamlessly edit photorealistic video from any input, including turning a person’s arm into mirror material mid-shot — it also published a careful, detailed post about expanding its content-verification tools.
Those are not two separate announcements. They are cause and effect, an hour apart.
Omni is, by design, a machine for producing convincing video of things that never happened. Google’s own showcase includes reimagining the action in footage you shot, dropping in characters who were never there, transforming a real moment into something else. It is genuinely, technically impressive. It is also the most capable consumer deepfake tool the company has ever put in front of the public, and pretending otherwise would be dishonest.
So Google did something I’ve rarely seen a company do so openly: it shipped the antidote in the same breath as the product. Every Omni video carries SynthID, Google’s invisible watermark. Verification for SynthID and for C2PA Content Credentials — the standard that shows whether media came from a camera or an AI — is expanding into Search, Chrome, and the Gemini app. Point at a clip, ask “Is this AI?”, get an answer. Google says the detector has already been used 50 million times.
That is responsible, and it deserves credit. It is also a quiet admission. A company does not build a lie detector for its flagship model unless it knows the model creates a problem big enough to need one. Google’s own cited research says it plainly: people correctly spot a high-quality deepfake only about a quarter of the time. The watermark isn’t a nice extra. It is structural.
Here’s the catch I’d want a reader to leave with. The verification works inside Google’s own surfaces — Search, Chrome, Gemini, Pixel cameras. The instant a video travels beyond that garden, the invisible watermark survives, but your ability to check it may not. Google is signing partners — OpenAI, ElevenLabs, Kakao, Nvidia, and Meta for Instagram — and that genuinely matters for how far trust can travel. But the honest near-term picture is this: whether you can verify a video may come down to which app you happen to be viewing it in. Authenticity is becoming a property of the platform, not of the file.
Strip away the stagecraft and I/O 2026 hands the ordinary gadget owner two real shifts, neither of which costs a cent in new hardware.
Your existing phone and laptop changed jobs. They are becoming terminals for agents that live in Google’s cloud — quietly more capable, and quietly more dependent on a connection, a compute-metered subscription, and an agent’s discretion while you’re not watching. The hardware story of 2026 isn’t the glasses in the fall. It’s that the devices already on your desk had their purpose rewritten without a firmware update.
And “is this real?” stopped being a question your eyes can answer. Google has given you tools to check, which is to its credit — and it has given everyone the tools to make the fakes, which is the cost. Verifying a video is now a deliberate act, like checking a lock, not something instinct does for free.
Neither of these is cause for panic. Both are cause for a different kind of attention. After more than a decade of these keynotes, the most useful thing I took out of this one wasn’t a product. It was two habits worth keeping: notice when an app quietly becomes an agent, and notice when you’re trusting a screen you haven’t actually verified.
The era of opening the app is ending. What replaces it is faster, genuinely smarter, and often a real convenience. It just asks for something the old era never did — that you trust what you cannot see.