Cloud Browsers Benchmark: Bypassing anti-bot protection

TL;DR: Dedicated stealth cloud browsers like bro, Browserbase and Browser Use handled the heat way better than open-source frameworks.

bro achieved the highest overall access rate at 83.50%, bypassing modern bot protections passively without relying on automated CAPTCHA solvers.

Hey there! If you're building scrapers or AI agents, you already know the struggle: most existing benchmarks focus on how well an agent clicks buttons or fills out forms. But let's be real - the biggest bottleneck is often just getting on the page without hitting CAPTCHA or getting blocked.

To figure out what actually works in the wild, we built an open-source browser benchmark and used a list of the 400 toughest sites protected by Cloudflare, DataDome, Akamai, PerimeterX and others.

We tested 9 of the top cloud browser providers and open-source frameworks to see who could get through and load the real content.

Benchmark results

We compared the success rates of each provider. A success means the browser successfully bypassed the protection and loaded the actual target data.

Automated CAPTCHA solving was deliberately disabled (where possible) to see how well these browsers bypass protections passively, using just their fingerprinting and proxy routing.

How we ran the tests

We kept the pipeline simple to make sure the results were fair and accurate.

1. Open URL: Hit the target URL.
2. Wait 60 Seconds: We gave the browser a full minute to ensure absolutely all page content loaded properly.
3. Capture Screenshot: Snap a full picture of the final state of the page.
4. Verification: Check if the protection was actually bypassed and the content was visible.

Also, to keep things fair, we used the exact same proxy network for bro, Obscura, Playwright and Selenium to ensure IP quality didn't affect their success rates.

The full benchmark results, including every screenshot we checked, can be found in our public spreadsheet.

Verification process

Verifying 400 sites isn't always straightforward. Some websites are sneaky - they'll load the header and footer perfectly, making a simple HTML check think the page loaded, while completely blocking the main content (like news articles or product listings).

To solve this, we used a dual-verification approach:

1. We fed all the screenshots to an LLM (Gemini 3.1 Pro) and asked it to judge if the target page content was loaded properly.
2. Then we manually verified everything the LLM labelled.

Why did we verify 4k images manually?

Because LLMs fail from time to time! There were tricky edge cases where the page partially loaded and the AI couldn't tell that the actual target data was missing. The difference between human and LLM in judging the success rate was around 1-3%.

Examples of a partially loaded page where the LLM struggled to label the result

Cost breakdown

We also tracked how much it costs to run these browser sessions.

~95% of the costs come from proxy bandwidth. To help keep your proxy usage low, bro allows you to customize proxy tier

Here's the bro session config

{
  "enable_proxy": true,
  "rotate_proxy": true,
  "proxy_tier": "basic",
  "proxy_policy": "full",
  "country": "US"
}

Some providers were tested on paid subscription tiers which naturally brings their rates down.

For example, Browserbase was tested on their $99/mo Startup plan (mid tier), while bro was tested on free Basic tier with zero subscription cost. If you're running bro's Ultra plan, usage rates drop by 50%.

What's next?

The anti-bot landscape is a constant cat-and-mouse game. Security providers update their detection algorithms daily, which means a browser that bypasses a protection today might get blocked tomorrow.

One of the most challenging site examples

To keep this benchmark accurate and relevant, we'll be running these tests periodically. We also plan to expand the dataset with more complex targets and include additional cloud browser providers and open-source solutions in future runs.

As for bro, our goal isn't just to maintain our current lead. We are constantly refining our stealth features to push our success rate even higher.

While achieving a perfect 100% pass rate is almost impossible in this space, we are determined to get as close as we can.

Source code & Dataset

We believe in transparency, so the source code for our evals is entirely public.

You can check it out on our GitHub repo: jsonifyco/browser-benchmarks.

FAQ