Elon posted that Grok 4.5 is “Opus-class” the same afternoon xAI pushed it live, and I spent about ten minutes poking at it before the obvious problem showed up. The model is genuinely good. It’s also stuck behind a chat box that has no idea what I’m doing in the other nineteen tabs.

That gap is the whole story here, and it has almost nothing to do with how smart Grok 4.5 actually is.

The app is a walled garden with a nice view

Open grok.com or the xAI app and you get a text field. You type, it answers. If you want it to work on something real, like a half-filled expense form or a Salesforce record or a thread in your inbox, you become the transport layer. You copy the page. You paste it in. You copy the answer back out. You do this maybe forty times a day and call it a workflow.

Grok 4.5 being Opus-class doesn’t fix any of that. A frontier model reading a screenshot you pasted is still just a model reading a screenshot you pasted. It can’t click the “next” button on a paginated table, it can’t see that you’re logged into the account the task actually lives in, and it definitely can’t tell that the invoice you’re asking about is sitting in the tab right next to it.

So where should a frontier model run?

In the browser. Where the work is.

A browser agent lives in your side panel and reads the page you already have open, using the session you already authenticated. Dassi does this as a Chrome extension, and the part that matters for a Grok launch is that it’s BYOK, bring your own key. You drop in your xAI key and now Grok 4.5 is looking at your actual DOM, not a description of it you had to assemble by hand.

Same weights. Same benchmark scores Elon was tweeting about. Completely different amount of useful context, because the model can finally see the thing you’re pointing at.

Nobody wants to be the copy-paste machine

I keep coming back to this because it’s the least glamorous problem in AI and somehow still the most common one. There’s a whole genre of productivity advice built around it, the prompt libraries and the browser bookmarklets that pre-fill your clipboard and the extensions whose entire job is to shovel page text into a chat window a little faster, and all of that infrastructure exists to paper over one design decision, which is that the model got put in an app instead of in the browser, and once you notice that decision you honestly can’t stop noticing it. When the model runs where the page already is, the copy-paste tax doesn’t get smaller. It goes to zero.

Why BYOK is the part that ages well

Here’s the durable reason to care, separate from any one launch.

Frontier models leapfrog each other now on roughly a six-week clock. Gemini 3.1 Pro, GPT-5.2, and now Grok 4.5, three “best model in the world” claims inside a single quarter, each one true for about as long as it took the next lab to ship. If your agent is welded to one provider’s app, every one of those launches is a migration project. You wait for that vendor to integrate the new thing, or you switch tools entirely.

With BYOK you just swap the key. Grok 4.5 comes out, you point your browser agent at it this afternoon, and if next month something beats it you point at that instead. The agent is the browser layer. The model is a setting. We wrote about why AI companies don’t trust each other and neither should you, and model portability is the practical version of that argument: don’t marry the model, date it.

And your key means your data goes to xAI directly, not through some middleman’s proxy that logs your prompts for “quality.” That’s a whole other reason BYOK wins, but I’ve ranted about proxies before.

The honest limitation

BYOK is not free money. You pay xAI per token instead of a flat monthly fee, so if you run huge jobs all day the metered bill can sting in a way a subscription doesn’t. For most people doing normal browser work, drafting replies, pulling data off pages, filling forms, it’s cheap. But I’d rather say it out loud than pretend the math always lands in your favor.

The other honest bit: a browser agent only helps with things that happen in a browser. Grok 4.5 in the xAI app is still fine for standalone questions where there’s no page involved. This isn’t a replacement for that. It’s the missing half.

The gap the demos never show

Every model launch ships with a highlight reel of the thing reasoning brilliantly about some puzzle. What the reel never shows is the model sitting three inches from the page you need help with, unable to reach it. Grok 4.5 is a real jump. Now go put it in the browser where you actually work, because a frontier model that can’t see your tabs is just a very expensive autocomplete.