Published
Updated
Browser Use wires Jev into the click loop: 7 seconds, $0.0039
This topic was created 31 days ago, and the information it contains may have evolved or changed since then.
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. On September 30, 2026 a desk loop on three routes never left the Google Flights start page, so date, price, and flight number were not confirmed there.
One of the first interactive Jev demos came from the browser-agent project Browser Use.
On September 17, Browser Use founder Gregor Zunic posted a clip of an agent searching for flights. He reported that the run took about seven seconds and cost $0.0039, with video playback at regular speed. The accompanying repository, jev-ultrafast, records the Zurich to London Google Flights path at 7.1 seconds.
At each step, the page converts into an indexed action space: click, type, select, scroll, wait, done, or blocked. Jev receives that table as state and selects both an operation and a target element in a single request. A small language model writes text only when the selected action is “type text”. Other actions remain direct selections from visible controls.
Operating a browser requires discrete actions rather than open-ended text generation. Zunic described updating the action space every step from DOM state, calling an LLM fallback only when text input is needed.
According to the repository, the Google Flights demo runs without site-specific scripts. The agent observes the page and executes actions toward a stated goal. Two decision heads share a single network round trip so the chosen target matches the chosen verb.
The seven-second duration and sub-cent cost are author-reported figures for this specific demo. Other builders shared similar loops using Stagehand or custom Chrome extensions, reporting costs near a tenth of a cent per task. Removing text generation from the execution loop reduced latency across each of these experiments.
On September 30, 2026 this desk did not replay jev-ultrafast. A separate Playwright loop, jev-1.13.0 for the click and gpt-6-luna for typed text, tried three one-way routes from the Google Flights start page: Zurich to London on November 12, San Francisco to New York on November 18, and Paris to Barcelona on December 3. Each try stayed on that start page. Misclicks were 8, 10, and 10. List cost for the Jev calls was $0.000487, $0.000489, and $0.000488. The gateway returned no dollar amount for the Luna tokens. Page reading on this machine was $0. The recordings show the explore page, not a fare. A constructed results URL for Zurich to London on November 12 did list easyJet at $57, British Airways at $87, and SWISS at $97, with “Nov 12” on the page, and it did not show a flight number. The agent did not open that URL.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Zunic's post and the jev-ultrafast README both describe a Zurich to London Google Flights path completed in about seven seconds. The repository states 7.1 seconds, while the post reports approximately seven seconds and $0.0039 with video shown at 1x speed. We did not replay the agent.
On September 30, 2026 this desk ran a different loop in headless system Chrome. It is not browser-use/jev-ultrafast and it is not Stagehand. Jev, response model jev-1.13.0, chose an operation and a visible control. gpt-6-luna was asked for text only when the operation was TYPE. The gateway returned tokens and no dollar field. Reading the DOM on this machine is the page-understanding step. Its price is $0. Each route was one try. A second try was reserved for a route that already showed a date, a price, and a flight number. None did.
Zurich to London on November 12, 2026: 43.2 seconds, 8 misclicks, 7 errors, 10 Jev calls, 11,607 input tokens, list cost $0.000487, Luna 638 input tokens and 111 output tokens, 80.5 ms spent reading the page. San Francisco to New York on November 18: 31.6 seconds, 10 misclicks, 0 errors, 11,650 input tokens, $0.000489, Luna 436 and 52. Paris to Barcelona on December 3: 35.9 seconds, 10 misclicks, 0 errors, 11,620 input tokens, $0.000488, Luna 637 and 49. All three ended on https://www.google.com/travel/flights?hl=en&curr=USD. The visible page was the explore view, with a map and a Search button. No date, no price, and no flight number were on it. Screen recordings from those three tries are webm files from the Playwright context. They show the same start page. They are not on this site.
A separate load, not the Jev loop, opened a Google Flights URL whose query was one-way Zurich to London on November 12, 2026. The title was "Zürich to London". The text included "Nov 12", "departing 2026-11-12", one way, 1 adult, and economy. The list included easyJet 8:15 PM to 9:00 PM, ZRH to LGW, nonstop, $57; British Airways 12:45 PM to 1:40 PM, ZRH to LHR, nonstop, $87; SWISS 3:30 PM to 4:20 PM, ZRH to LHR, nonstop, $97. No flight number of the form "BA 712" was in that text, and clicking the easyJet line did not add one. Those fares are what the page showed that afternoon. They were not booked, and the agent did not reach them. The 7.1 seconds and $0.0039 remain the repository's figures.
The claim of no site-specific scripts is from the repository author. The dual-head design (operation and target) is described in the repository rather than independently verified by us.
Compare
Kyle Jeong's Stagehand test reported $0.001 and near-instant response. RaZaan's Chrome extension routed through Vercel's gateway. The Tetris demos apply the same pattern across a finite move list, reporting 300 to 400 ms per decision.
None of these represents a standardized computer-use evaluation. They share a consistent architecture: Jev selects from visible page controls, and a language model provides text only when a field requires input.
Terms
- Action space
- The indexed list of valid page actions for the current step (click, type, select, scroll, wait, done, blocked).
- Target element
- The specific page element receiving the selected action. The repository states that both heads share a request to avoid mismatched actions and targets.
- LLM fallback
- A small language model used exclusively when the chosen operation requires text input.