A hundred announcements in two hours. Gemini 3.5 Flash, Gemini Omni, Antigravity 2.0, Spark, Daily Brief, Ask YouTube, Project Genie, a personal agent on macOS, a Street-View-to-simulator thing I never thought I'd have to type. I lost count somewhere in the second hour. So did most of the room. So did, almost certainly, most of the people who wrote the recaps.

Five things worth keeping. Five things you can safely ignore. Read the rest of the timeline next week if you fancy. This is the spine of it.

The signal.

Signal · 01

Gemini Omni. One model, every modality, every surface.

Text, image, video, audio in. Anything out. Rolled straight into Shorts, into Create, into the Gemini app, into Workspace. You can talk to it like a person, on the surface you already use, and it answers back in whichever format the moment needs.

The model became a colleague. Quietly.

Signal · 02

2.5 billion. 900 million. 3.2 quadrillion.

AI Overviews at 2.5B users. AI Mode past 1B on its own. Gemini app 400M to 900M in twelve months. 3.2 quadrillion tokens a month across Google, up sevenfold. A third of the planet is already inside the new interface and most of them haven't noticed. Bonkers!

The only stat that should have mattered all week.

Signal · 03

The surface bet.

Search is a conversation. Chrome is a runtime. Workspace is an agent harness. Android has a built-in personal agent. Same model under all of them. The surface is the moat.

Owning the surface owns the morning routine.

Signal · 04

Antigravity 2.0 and the agent-first dev platform.

Agents move out of the chat box and into actual workflows. If you're shipping product right now, this is the announcement that quietly resets the floor under everything you build for the next two years. Optimise for it.

The boring announcement that runs the next decade.

Signal · 05

Ask YouTube and the Daily Brief.

The chat panel is the new homepage. You don't click the video. You don't open the publisher's newsletter. You don't scan the ten blue links. You ask. Brutal for the open web. Real for everyone else.

If your work only lives as a clickable link, you're betting on a behaviour in decline.

The noise.

Noise · 01

The benchmark wars.

Is Flash faster than GPT? Does Omni beat Veo? The leaderboard moves every six months. The surface moves every five years. Stop watching the wrong scoreboard.

Real metric, wrong horizon.

Noise · 02

Project Genie and the simulator demo.

Astonishing. Honestly. Walk through a Street View location as a 3D simulator and it does feel like the future. But there's no consumer product fit yet, no obvious workflow, no business model. Cool, parked.

Bookmark it. Don't theorise on it.

Noise · 03

The four-thinking-levels marketing.

Useful if you're an engineer wiring an inference budget. Useless to anyone else. It's a knob dressed up as a paradigm. Move on.

A toggle is not a thesis.

Noise · 04

The hundred-announcement count itself.

Stratechery called it I/O Spaghetti and he's right. A hundred things is what you ship when you don't quite trust any one of them to carry the room. Grade Google on what's still being talked about in October, not on what got listed on Monday.

Volume is a tell. Usually a nervous one.

Noise · 05

Whatever the timeline is screaming about by Wednesday.

Yes, the SynthID watermark debate. Yes, the next "is this AGI" thread. Important conversations. None of them the I/O story. Mute them and come back to the surface bet.

If it didn't change the surface, it didn't change anything.

The thesis, in a line.

The model is the demo. The surface is the moat.

Veo has a great model. LTX has a great model. Pika, Runway, all of them. I use them. I love them. They are doing astonishing work. But none of them has an OS. None of them has a browser. None of them has the search bar that you, your mother, and the bloke who installed our boiler all reach for when they want to know something. Distribution used to be a slow, late-game advantage. Now it's the runtime. The surface you're on is the agent. The agent is the model.

Google did one thing on Monday that nobody else can do. They launched a frontier model and, in the same breath, deployed it into the front door of the internet for 2.5 billion people. That asymmetry might be the most important one in the industry right now!

I wrote a piece a few weeks back called Peers, Not Seats, arguing the per-seat SaaS model is over because the seats are being absorbed by agents. I/O 2026 is the other side of that argument. The seat is being absorbed. The browser tab the seat used to sit inside is being absorbed too. The surface itself is the product now. And the company that owns the most-used surface on Earth just stapled a frontier model to the back of it.

Standalone models vs. surface-and-model stack The standalone model Lab + API + user goes to it Frontier model API + chat surface User · opens it deliberately Distribution: opt-in. User has to remember to come. Moat: model quality. Lasts ~6 months between leapfrogs. The surface + the model User is already inside it Search + Chrome + Workspace + Android Gemini app + AI Mode + AI Overviews Frontier model (Gemini 3.5 Flash / Omni) Daily Brief · Ask YouTube · Spark · Antigravity Distribution: ambient. 2.5B users already on the surface. Moat: surface ownership. Resets only when habits reset.

Fig. 1 · Two shapes. One has to be opened. The other is already open.

The honest counter-reads.

Two outside reads made me uncomfortable. Which is usually the tell.

Ben Thompson at Stratechery called it I/O Spaghetti... a hundred things is what you ship when you don't trust any one of them to carry the room, and he's right about that. He also asked, pointedly, whether DeepMind is actually aligned with Google's commercial business or bolted to the side of it. Fair question. The seams will tell us something eventually.

Casey Newton at Platformer asked the harder one. Are we summarising the open web to death? Ask YouTube, Daily Brief, AI Overviews... each one a small convenience for the user and a small wound for the people who make the thing being summarised. Stack them up and the chat panel is quietly eating the click-through economy alive. Not hypothetically. That's what they showed us on Monday, with a smile.

I make stuff for the web. I read stuff on the web. I've watched publisher friends try to find a business model in a world where the click-through is being eaten by the chat, and most of them are struggling. The writers, designers, makers, and brands who survive the next five years are the ones who build for the agent first and the SERP second. If your work only lives as a clickable link, you're betting on a behaviour that's already in decline. I'm pretty bloody sure which side I want to be on.

The bit nobody put on a slide.

Sometime this week Lucas, who's ten, asked Ask YouTube to give him the best bits of the I/O keynote instead of watching the whole thing. Straight-faced. No idea that wasn't an obvious thing to do.

Shit! He was right.

I keep getting hung up on what's new about agents, what's new about the brief replacing the prompt, what's new about the AI as a colleague, a co-creator, a partner-in-the-room. For Lucas and Olivia none of it is new. The interface always answered them. The machine always thought and did. They don't see the profound bit because there isn't a before-shape to compare it to. It's water.

Which is, honestly, the whole point. We're not arriving somewhere strange. The kids are already there. The grown-ups are the ones still squinting at the keynote.

Si