Published
Updated
Stagehand, Chrome, and a11y trees: computer use as a Jev question
This topic was created 31 days ago, and the information it contains may have evolved or changed since then.
Kyle Jeong of Browserbase and other builders structured computer-use around structured observation, typed queries, and automated execution, reporting sub-cent task costs. On September 30, 2026 the same task reached the Jevons paradox article on Wikipedia and stopped short on DuckDuckGo and Britannica.
Following Browser Use’s flight search demonstration, developers applied similar patterns to other browser automation tools: capturing page state, submitting valid actions as discrete questions to Jev, and running deterministic execution scripts.
Kyle Jeong of Browserbase demonstrated a remote-browser run pairing Jev with Stagehand, reporting an execution cost of $0.001. The process captures the page accessibility tree as state, submits possible UI actions as questions, and executes the selected action via Stagehand without generating open-ended plans or running vision models on screenshots.
Joey Kudish published a related project, jev-browser, reporting lower latency and costs compared to conventional language models across tested browser tasks. Separately, RaZaan demonstrated a Chrome extension connecting to Jev through Vercel AI Gateway to drive live browser sessions.
These implementations restrict decisions to an explicit action space. Limiting selection to currently active controls prevents models from inventing nonexistent element selectors. Text input still requires an auxiliary language model when filling form fields, though non-typing actions comprise the majority of interactions across many websites.
These experiments are early demonstrations rather than standardized benchmarks against established computer-use agents. They show developers applying Jev to discrete interaction steps, reserving conversational models for open-ended text input.
On September 30, 2026 this desk ran that shape of loop itself, in Playwright, on Wikipedia, DuckDuckGo, and Britannica. The task was the Jevons paradox article. Stagehand was not installed. Wikipedia ended on the article in 7.9 seconds, with no misclick, Jev list cost $0.000052, and a Luna call of 58 input tokens whose dollar price the gateway did not return. DuckDuckGo accepted the query and then showed a duck-square challenge the control list did not include. Seven Search clicks timed out. Britannica stayed on a Cloudflare check, and Jev returned BLOCKED. Page reading on this machine cost $0. The three Jev bills were $0.000052, $0.000406, and $0.000024.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Jeong's reported $0.001 cost and near-instant execution reflect his post rather than an official Browserbase benchmark. Kudish's jev-browser repository exists, implementing an accessibility snapshot, question, and execution loop.
RaZaan's clip demonstrates a Chrome extension using Vercel AI Gateway, matching the public typesafe-ai/jev gateway listing. None of these demonstrations represents a controlled benchmark. We did not run Stagehand with Jev.
On September 30, 2026 this desk ran one task in headless system Chrome with Playwright. The task was to search for the Jevons paradox and stop on the article that defines it, not on a list of results. Jev, response model jev-1.13.0, chose the operation and the control. gpt-6-luna supplied text only for TYPE. The gateway returned no dollar field. Local DOM reading is the page-understanding cost, at $0. This is not Browserbase and not jev-browser.
Wikipedia, from the main page: verified. The final URL was https://en.wikipedia.org/wiki/Jevons_paradox. The title contained Jevons, and the page text contained efficiency together with consumption, demand, or coal. Wall time 7.9 seconds, 0 misclicks, 0 errors. One TYPE of "Jevons paradox", then the article was already open. Jev used 1,240 input tokens and 292 output tokens, list cost $0.000052. Luna used 58 input tokens and 8 output tokens. DOM reading took 24.5 ms.
DuckDuckGo: the query "Jevons paradox definition" was submitted, and the URL carried that query. The run did not open an article. The screenshot shows a challenge dialog, "Select all squares containing a duck." Those squares were not in the control list the model saw. The next seven actions were clicks on the Search button. Each click timed out. Wall time 50 seconds, 7 misclicks, 7 errors. Jev used 9,674 input tokens, list cost $0.000406. Luna used 61 input tokens and 9 output tokens. DOM reading took 72.7 ms. The definition was not checked.
Britannica: the title stayed "Just a moment..." after a 12 second wait. The page was Cloudflare's bot check. Jev chose BLOCKED. Wall time 37 seconds, 0 misclicks. Jev used 562 input tokens, list cost $0.000024. Luna was not called. DOM reading took 7.1 ms. No article text was read. Jeong's $0.001 figure remains his post.
Compare
Browser Use's flight search demo uses a similar approach on a different harness (reporting 7.1 seconds and $0.0039). The Tetris demos share this structure with an explicitly bounded action space. Vercel provided the initial public gateway integration, followed by Cloudflare, Netlify, and OpenRouter.
Accessibility trees supply a more compact text input than raw screenshots, enabling these loops to avoid continuous image captioning models.
Terms
- Stagehand
- Browserbase's library for automating remote browser sessions through structured interaction steps.
- Accessibility tree
- The structural snapshot of a webpage exposing element roles, labels, and states. These workflows supply this tree to Jev as state rather than transmitting screenshot images.
- AI Gateway
- Vercel's managed inference proxy, through which RaZaan accessed typesafe-ai/jev without configuring direct TypeSafe account keys.