Did you run the other browsers yourselves?

No. Every row except ours is Browser Use's own published runs, read from their repository at commit 661ce392, with their agent driving each provider's browser. We added our 3 runs and scored all of them with their code, unchanged. Scored that way, their runs give the figures in their own write-up.

Why 71 tasks and not 80?

Browser Use's scoring code leaves out the hCaptcha, GeeTest and Temu tasks and one site with no protection on it, so every run is scored on 71. We used the same code so the rows compare. On all 80 tasks, our 3 runs reached 92.1%.

What are the two plain browsers?

Browser Use's controls: an ordinary local browser on a datacenter IP, run without any service around it, to show the sites are hard. Headless, it reached 2.2%; with a visible window, 49.3%.

Is the agent the same?

No, and that is the main difference. Browser Use drove every other browser with their own agent; we drove ours with ours. The judge only asks whether the page was reached, not how well the task was done, which keeps the score about the page rather than the agent.

Yes. The tasks, the judge and the scoring code are open in Browser Use's benchmark repository, with the steps to run it. The task list is encrypted there to keep it out of training data.

outcrawl

Benchmark

Your agent gets the page. We measured it.

Browser Use built a test from the 80 sites where AI agents got blocked most in their own traffic. We ran it on Outcrawl and scored it with their code.

Get a free API key How it was measured

Run on 7 October 2026 with their scoring code, unchanged.

Pages reached

Outcrawl

92.5%

Mean of 3 runs on 71 scored tasks.

Next best, Browser Use Cloud: 81.4%
95% interval: 91.5% to 94.4%

Results

Every browser, on the same test. Three ways to read it.

Outcrawl’s row is our 3 runs. Every other row is Browser Use’s own published runs: their agent on each provider’s browser. Same tasks, same judge, same scoring.

Outcrawl92.5%
Browser Use Cloud81.4%
Anchor76.8%
Kernel67.1%
Browserless54.0%
Chromium, headful (control)49.3%
Steel47.2%
Browserbase41.9%
Hyperbrowser39.9%
Chromium, headless (control)2.2%

Pages reached by each provider, split by the security vendor in front of each site
Provider	Cloudflare	PerimeterX	DataDome	Akamai	reCAPTCHA	Others
Outcrawl	88%	96%	90%	100%	100%	78%
Browser Use Cloud	93%	81%	69%	85%	80%	40%
Anchor	78%	81%	77%	69%	83%	50%
Kernel	81%	31%	62%	100%	89%	67%
Browserless	59%	28%	69%	67%	83%	11%
Steel	59%	28%	54%	31%	67%	50%
Browserbase	67%	22%	12%	31%	83%	42%
Hyperbrowser	43%	30%	36%	25%	83%	44%

Others is Kasada, Shape and one site-built check.

Method

How it was run and scored. Their test and their code. Our runs.

01
Short tasks, signed out.
On each site the agent searches, opens a page and reads a link. No sign-in, so what is tested is whether the page loads for the agent.
02
One question for the judge.
Browser Use’s AI judge reads each run and decides only whether the agent reached the page. A page that never loads counts as not reached.
03
Three full runs, on production.
All 80 tasks, 3 times, on 7 October 2026, through Outcrawl’s agent on the same API your key calls. A task that errored before it ran would count as failed; none did.
04
Their scoring code, unchanged.
Run over their published runs and ours together. It scores 71 of the 80 tasks, takes the mean over runs and adds a 95% interval.

Browser Use’s repository at 661ce392 Browser Use’s own write-up

mean 92.5%

90%91%92%93%94%95%96%

95% interval

Run 165 of 71
Run 265 of 71
Run 367 of 71

What the interval means: resampling our 3 runs, the mean lands in this range 95 times in 100. It shows how much the result moved from run to run, not how any browser does on other sites.

Limits

What this does not show. Read it with these in mind.

The agent is not the same. Browser Use’s agent drove every other browser; ours drove Outcrawl. The judge asks only whether the page was reached, which narrows that difference without removing it.
The dates are not the same. Browser Use published their runs in March 2026; ours ran on 7 October 2026, and sites change. Two of the Cloudflare sites, Copilot and Perplexity, now ask every signed-out visitor to sign in; a plain browser on the same network got the same that day.
Cost is not compared. Their published runs record the model’s cost but not the browser’s, and time per page includes the agent’s own time.
It is 71 of the 80 sites. Their scoring leaves 9 out (see the questions below); on all 80, our runs reached 92.1%.

FAQ