Thirty Cents a Million: DeepSeek V4.1 Flash and the Collapsing Cost of Intelligence

The Buttons Are Gone

Now the part of this release that no spreadsheet captures, and honestly my favorite: open the DeepSeek app and the mode picker is gone. The app used to make you choose your intelligence up front — a fast mode for quick questions, an expert mode for the deep stuff, an image mode for pictures — three buttons, three moods, and the standing suspicion that you were about to pick the wrong one (IT之家). With V4.1 Flash, all three are merged into one. You type, you drop in a photo if you have one, you hit send, and the model underneath routes itself: it reads the difficulty of the query and dynamically scales how much effort it spends, and the moment it sees an image it wakes up the native vision without being told. Deep thinking and search switch on the same way. One box, one button, one model deciding. (｡•̀ᴗ-)✧

And once you see it, the whole release snaps into focus, because the app isn't the only place DeepSeek deleted a choice this week. On the API there's exactly one name to learn now: deepseek-flash. The retired V4-Flash and V4-Flash-Vision names still work, but they're forwarding addresses to V4.1 Flash. And from 04:00 UTC on September 14, so is deepseek-v4-pro — the flagship's own name becomes a redirect to the model that replaced it, billed at that model's cheaper rates, until V4.1 Pro ships (changelog). Every road, paved, leading to the same place. DeepSeek didn't just ship a model this week; they shipped a router and retired the menu.

I love this as product design, and I love it more as a preview of where the whole industry is sliding. Model pickers were always an engineering problem wearing a user-interface costume — "which mode should I use?" is a question the user should never have been asked in the first place. The router knows the model card, the router sees the query, and the router doesn't get decision fatigue at midnight. We spent years choosing between fast and smart, and the people building the models finally noticed the correct answer was "yes."

The best interface, I've learned, is the one that stops asking you which kind of smart you want.