Someone reposted a screenshot on Saturday with the headline “Kimi: Threat or menace?” and I actually laughed, because it’s the exact tone the whole feed has taken this week. Moonshot’s model is cheap, it’s fast, the weights are out in the open, and everyone has already decided it’s either going to eat everyone’s lunch or it’s benchmark bait with good PR. The arguments are fun. Most of them are also completely untethered from anyone having run the thing against their own work.

So I ran it on mine.

The fight is about the model. It shouldn’t be about the app.

Something bugs me about model launches. Every time a new one drops, the assumption is that trying it means adopting a new destination: a new chat site, a new desktop app, a new tab you’ll forget about by Thursday. You go there, you paste some text, you judge it inside a sandbox that looks nothing like your actual day.

But the model is just an API. The interesting question was never “is Kimi good in Kimi’s own playground.” It’s whether it’s any good at the stuff you’re already doing, in the places you already do it, against pages you’re already logged into. I made basically this same point about GPT-5.2 in why I stopped opening the chat site at all, and it holds for whatever model is trending this week too.

That’s what BYOK is for. You bring your own key, you point a browser agent at it, and the model runs where your work lives instead of where the vendor wants you to visit.

Drop the key in, keep your tabs

The setup is dull, which is the point. I have Dassi in my side panel, I opened settings, pasted a Moonshot API key, picked Kimi from the model list. Done. No new browser, no migration, no re-logging-into anything.

And because Dassi works inside the Chrome tab I already have open, Kimi inherited every session I was already in. My Gmail. A half-filled vendor form. Three research tabs I’d been ignoring. The model didn’t need OAuth scopes or a connector or an integration. It could just see the page, the same way I could, because it was running in my browser and not on some rented server in a datacenter that has never met a single one of my login cookies.

What it did

I gave it boring jobs on purpose. Boring jobs are the honest test.

First, triage: read the top of my inbox, tell me which threads actually needed a reply today, draft the two that did. Kimi was fast. Noticeably fast, the way you only notice when you’re used to waiting. The drafts were fine. Not brilliant, not embarrassing. I fixed a sentence in each and sent them.

Then I aimed it at a pricing page that buries its plan comparison behind a row of tabs, and asked it to pull the numbers into a table. It clicked through, read each panel, gave me the table. One number was wrong, which I only caught because I double-checked. So double-check.

Was it better than GPT-5.2 at this? No, honestly. About the same, on these tasks. But it cost a fraction of what I’d normally spend, and for inbox triage and scraping figures off a page, “about the same for way less money” is the entire ballgame.

Fine is a result

“Fine” is a verdict people skip past because it isn’t dramatic. But fine, at a tenth of the price, running against my real tabs, is exactly the outcome that changes what I actually do on a Tuesday.

Cheap changes what you’re willing to run

There’s a threshold thing that happens with price. When each little automated task costs real money, you ration them without really deciding to: you fire the agent for the big stuff, the reports and the multi-step slogs, and the small daily annoyances stay manual because none of them individually clears the bar of being worth the spend. Drop the cost by an order of magnitude and that quiet math flips over. Suddenly it’s fine to have the thing skim every newsletter, re-label a pile of receipts, pull one figure off a dashboard you glance at twice a day, summarize a thread you don’t have time to read. The tasks didn’t get more valuable. They just stopped costing enough to think about, and once you stop thinking about the cost you start handing over the work you used to grind through by hand out of sheer stubbornness. Cheap models don’t only do the same work for less; they quietly widen what counts as worth automating in the first place, and that second effect is the one people keep underrating when they line up to argue about which benchmark Kimi did or didn’t top. The benchmark fight is loud. The pricing shift is the thing that’ll actually change your afternoon.

The part nobody’s arguing about

If Kimi turns out to be overhyped, you swap the key back and you’ve lost nothing but ten minutes. That’s the whole case for keeping the model separate from the tool. I wrote about why BYOK stops being optional once you notice how much of your data these vendors would love to hold onto, and cheap open challengers only make that argument louder.

The threat-or-menace headlines will keep coming. Every month has a new one now. What doesn’t change is the browser you’re reading them in, which is still the most useful place to find out whether any of it is real. You can add Dassi to Chrome, paste whatever key is trending this week, and let the tab you’re already in settle the argument.