Browser Use wires Jev into the click loop — 7 seconds, $0.0039
Gregor Zunic showed a 1x-speed Google Flights run where Jev picks the action and DOM target each step, and a small LLM types only when it has to.
The first Jev demo that felt like a product, rather than a slide, came from the browser-agent world.
On September 17, Browser Use founder Gregor Zunic posted a clip of an agent finding flights. He said the run took about seven seconds and cost $0.0039. The video, he noted, is real time — not sped up. The repository behind it, jev-ultrafast, describes the Zürich → London Google Flights path as 7.1 seconds.
The architecture is the interesting part. Each step, the page is turned into an indexed action space: click, type, select, scroll, wait, done, blocked. Jev gets that table as state and, in one request, chooses an operation and a target element. Only when the operation is “type text” does a small language model write the characters. Everything else is a multiple-choice question over what is actually on the screen.
That split is why people on X reached for the same metaphor: using a chat model to operate a webpage is a long essay about which button to press. Clicking is a discrete choice. Jev’s interface matches the choice. Zunic’s notes mention a new action space every step, DOM state, and an LLM fallback only for typing.
The repo is explicit that there are no site-specific scripts for the Flights example. You give a goal; the loop observes, decides, acts. Two decision heads share one network round trip so the chosen target cannot belong to the wrong verb.
Treat the 7-second / sub-cent figures as author-reported measurements of this demo, not a universal browser-use benchmark. Other posts the same day showed similar loops with Stagehand or custom Chrome extensions, at claimed costs around a tenth of a cent per task. The pattern is consistent even when the exact number moves: keep generation off the hot path, and the browser stops feeling like a slideshow.