We benchmarked Yad against browser-use, then checked the answers
Twelve identical tasks, free models on both sides, no cherry-picking. Yad came out faster and used fewer steps. The finding that actually mattered wasn't in that table, it was what turned up when we went back and checked whether the answers were even true.
We picked browser-use because it's the open-source agent people already compare Yad against, and we picked two public scraping-practice sites, books.toscrape.com and quotes.toscrape.com, built specifically for this kind of exercise, not a production app under real load. Twelve tasks: find a price, count results, navigate into a category, extract a list, that range. Free-tier models on both sides, same tasks, same grading script. Small sample, one site pair. We're not calling this a final verdict, just the first time we actually measured it instead of assuming it.
The numbers
Yad was faster on 11 of the 12 tasks and used less than half the steps on average. The one task browser-use won on time, Yad hit a genuine unexpected page navigation mid-task, recovered on its own, and still landed the correct answer, just slower than a clean run.
The answer that wasn't true
Task seven asked both agents to find a five-star-rated book on the site's homepage and report the title and price. Yad came back with a partial: star ratings on that page are rendered as small icon classes in the HTML, not as plain text, so it said it couldn't confirm one from a text read alone and offered to dig further. browser-use came back confident: "A Light in the Attic, five stars, £51.77."
We checked the site's actual HTML. A Light in the Attic is a three-star book. The real five-star books on that page are Sapiens: A Brief History of Humankind, Set Me Free, Scott Pilgrim's Precious Little Life, and Rip It Up and Start Again. Both runs used text-only extraction, no vision, so browser-use wasn't looking at the page any differently than Yad was. It didn't find the rating either. It stated a guess as fact, and our own grading script counted it as a pass anyway, because the script only checked for a price symbol in the answer, not whether the specific claim was actually correct.
What this actually means
This isn't really a browser-use-specific problem, it's a property of confident-sounding language models generally: an agent that always returns a clean, fully-formed answer looks more finished than one that sometimes says "I can't verify this," even when the first one is wrong and the second one is right. Our own scoring script had the exact same blind spot until we sat down and read the raw answer text by hand instead of trusting the pass/fail flag.
Two things we're keeping from this. First, a benchmark score is only as honest as its grading, checking for the presence of a keyword isn't the same as checking a claim is true, and we're tightening our own scoring for exactly that. Second, an agent declining to answer when it genuinely can't verify something isn't a worse result than a confident guess, it's a better one, and that's something we now want to measure directly instead of only measuring speed and step count.
Twelve tasks, one site pair, free-tier models on both sides. This isn't a leaderboard claim. It's the first time we actually checked.