Google announced the Googlebook at $899 and the whole pitch is Gemini. I read the spec page twice trying to find the part where the hardware actually changes what the AI can do for me, and I came up empty. The device is, functionally, a browser in a shell. And the AI on it runs beside the browser, not inside it.

Which means it starts every task as a stranger.

A laptop whose main app is Chrome

Chromebooks were always honest about being a browser with a battery. That was the appeal: cheap, fast to wipe, nothing to maintain. The Googlebook keeps that shape and adds a Gemini key and some NPU headroom, positioned as the reason to spend $899 instead of $250 on a used Chromebook that already opens the same tabs.

But the model isn’t running on that NPU for anything interesting. Gemini’s real work happens in a datacenter, same as it did yesterday, same as it does on the phone in your pocket.

So what does Gemini actually get handed?

A prompt, some page text if you’ve invoked it on a page, and whatever Google’s own apps volunteer through Workspace.

That’s it. If you ask it to do something in Salesforce, it doesn’t have Salesforce. It has a description of Salesforce from its training data and a willingness to guess.

The Salesforce version of this

Here is the comparison I keep running in my head, and it has nothing to do with benchmarks.

I’m looking at an opportunity record in Salesforce. I’m logged in through Okta, my session has been alive since Monday morning, and the record shows a close date that’s wrong because the customer moved their fiscal quarter. I want the AI to read the last three activity notes, update the close date, and draft the internal Slack message explaining why.

A cloud model on a new laptop gets: nothing. It cannot see the record. If you tell it to go look, it either asks you to paste the page or it spins up a fresh browser session somewhere in us-central1 that hits my company’s SSO wall and stops. Okta doesn’t care that Gemini is very smart. Okta cares about a cookie Gemini doesn’t have.

A browser-native agent running as an extension in the Chrome profile I’m already using gets: the DOM of that live tab. The actual rendered record, including the fields that only appear after the page’s JavaScript finishes, including the activity timeline that loads lazily, including the account owner’s name spelled the way it’s spelled in our org rather than the way the model would assume. It acts through my session because it is my session. No OAuth grant, no service account, no admin ticket to approve a new integration.

That’s the gap. It isn’t a model gap, and it will not close with a faster chip. We wrote about this when Nvidia was pitching $200B worth of AI agent PCs, and the Googlebook is the same bet wearing a $899 price tag instead of a keynote.

Where dassi sits in this

dassi is a Chrome extension. It lives in the side panel of the browser you already have open, reads the page you’re already looking at, and clicks through your existing login state. You bring your own model, including Gemini if that’s the one you like, so the intelligence is identical to what the Googlebook ships with. The plumbing is what’s different. It’s on the Chrome Web Store, free, and it works on the laptop you’re reading this on.

Limits, because an extension is not magic

An extension only sees tabs you open it in. That’s a real constraint and I’d rather say it than have you discover it on a Tuesday.

It doesn’t crawl your browsing history. It doesn’t sit in the background reading the forty tabs you left from last week’s research spiral. If the Gmail tab isn’t the tab you’re in, it isn’t context. You point it at something, it works there. Some people find that annoying and want an agent that has ambient access to everything, and I understand the impulse, but I think scoping the agent to the tab in front of you is the correct default and not a compromise — an AI with permanent read access to every authenticated session in your browser is a hell of a thing to hand a language model that can be talked into stuff by a malicious page.

There’s also plenty an extension can’t reach. Native apps, your filesystem, anything outside Chrome. If your work lives in Excel on a desktop, a browser agent is the wrong tool and I won’t pretend otherwise. We went through the full boundary in what a Chrome extension agent can actually see.

$899 for a context problem

The Googlebook will probably sell fine. Chromebooks sell to schools and to people who want a laptop that can’t get infected, and a Gemini key is a reasonable thing to put on one.

Just notice what you’re buying. You are not buying an AI that can do your job in the browser, because the AI is still outside the browser, looking in through a window it can’t open. The cookie jar is already on your desk.