How we test antidetect browsers

• One Linux x86_64 server with 16 cores, 31 GB RAM, and no GPU. The display is a virtual X display (Xvfb). Every product has the same host.
• Free plans or free trials only, with no payment. A product with paid plans only is listed as not tested. So is a product with no Linux desktop app.
• Identity: Windows, the newest browser version that the product offers, and US English. Timezone, geolocation, and WebRTC stay at the product default, which is usually “based on IP”. All other settings stay at the product defaults. No tuning.
• Two profiles per product, A and B, because the smallest free plan allows two.
• We wrote down the app version and the engine version of every product, and turned off automatic updates where the app allows it.
• AppImages were extracted and run, because this host cannot mount them.

Before each product, stock Google Chrome for Linux, with no changes, ran the real anti-bot checks on the IP of that product. If the control failed a check, the IP was replaced once. If the control failed again, that check would be dropped for that product and its points spread over the other real checks.

With the final IPs, stock Chrome passed the first Cloudflare page, scored 0.9 on reCAPTCHA v3, and had no Fingerprint Pro flag, for every product. So a flag on a product comes from the product, not from its IP.

• Each product got a fresh US ISP proxy of its own, from the same provider. Both profiles of a product used that IP.
• At least 5 minutes passed between two DataDome page loads on one IP.
• The first Cloudflare page sets a clearance cookie that lasts a year. So the second and third Cloudflare tries used two other challenge pages on other domains, which no profile had visited.

• Each profile started the way the product’s own app starts it.
• We opened every page by typing its URL with keyboard input from the operating system. We attached no CDP or Playwright session to any browser that an app manages.
• Camoufox and CloakBrowser are libraries, so they started through their own documented API. That is their only normal use.
• AliasMode profiles started through the AliasMode app’s own launcher.
• We never clicked a Cloudflare checkbox.
• Evidence: a screenshot for every verdict, the page text copied with the keyboard (Ctrl+A, Ctrl+C), the probe data, and a screen recording of each run.

The probe page hashes each of these signals in profile A and in profile B of the same product. A signal counts when the two hashes differ. A value that is empty in both profiles counts as the same value.

• Canvas 2D image
• WebGL image readback
• Audio (an offline render)
• Text metrics
• DOMRect and emoji size
• Font list
• Speech voices

Each value counts once. They are checked against stock Chrome on Windows, or stock Firefox on Windows for Firefox engines.

1. navigator.platform is not Win32.
2. The user-agent client hints do not say Windows, or the platform version is impossible.
3. The user agent, the client-hint version, and the real engine version disagree.
4. userAgentData is missing, or its brand list is malformed (Chromium engines).
5. Core Windows fonts are missing, or Linux-only fonts are visible.
6. The WebGL renderer is not a Windows ANGLE or Direct3D string, or a software renderer (SwiftShader or llvmpipe) is visible.
7. The timezone differs from the timezone of the proxy IP.
8. navigator.languages and the Accept-Language header disagree.
9. WebRTC shows a public IP that is not the proxy IP.
10. The speech voices are not Windows voices.
11. Screen values are impossible: the available size is larger than the screen, or the offsets or the pixel ratio are odd.
12. hardwareConcurrency or deviceMemory has an impossible value.
13. navigator.webdriver is true, or automation traces are visible.
14. Native-code checks fail on patched functions (toString or prototype lies).
15. Any other value that stock Chrome or Firefox on Windows never shows, written down with the evidence.

Launch time runs from the start command until the probe page is fully loaded. RAM is the total proportional set size (PSS) of the browser process tree, with one probe tab, 60 seconds after the start. Each product had a warm-up launch, and then 3 measured launches. The pages show the median of the 3.

• DataDome (the GitHub sign-up page): stock Chrome was blocked on every IP where we tried it, 12 in all. So the host or its IP range is flagged, and the check could not separate the browsers. Its 10 points went to the other real checks.
• iphey: it gave no verdict, even for stock Chrome on this host. Its points went to CreepJS, BrowserScan, and Pixelscan.
• Pixelscan often stayed on “scanning” or “Collecting Data”, even for stock Chrome. Only a verdict that the page showed earns points.

• Cloudflare is scored on 3 pages instead of 1 try. This changed on 2026-10-29, after the results, for every browser. One try proved too random: a retry of the same page 15 seconds later could change the result. Stock Chrome passed each new page on one of the test IPs before we used it.
• Cloudflare page 3, Dolphin Anty and Octo Browser: the click that our harness uses to copy the page text opened a link on the page before the pass text was read. The try was repeated once in the same profile, and it passed. The profile may already have held a clearance, so these two passes count, with a flag in our notes.
• CreepJS lies were counted from the saved page, the same way for every browser.
• Timezone: the geo databases disagree for these IPs (Texas or Michigan). A timezone that matches ipinfo, ip-api, or the IP timezone that BrowserScan shows counts as correct.
• This host blocks the Chrome sandbox, and we have no root. So every Chromium browser ran with --no-sandbox. The warning bar that this adds is not counted as an impossible value.
• CloakBrowser (free 146) and MoreLogin Chrome got a new IP once, because stock Chrome failed Cloudflare on their first IP.
• Octo Browser has no free plan, so it ran on its free 2-day trial (code TRIAL, no payment). GoLogin ran on its free 7-day trial.
• The Octo Browser app clears the library path of the host, so its engine got a small wrapper that restores it. The wrapper adds a few milliseconds to the launch time.
• Dolphin Anty refused one temporary email domain and accepted another. Its new profiles default to Linux, so we chose Windows 11.
• GoLogin’s new profiles first launched Orbita 109, until the app’s own “Update profiles” step. The Orbita 154 run is scored.
• Donut Browser’s free plan offers no Windows identity, so it ran with a Linux identity. Its impossible values are counted against Windows, as the method requires.
• The AdsPower free plan holds 2 profiles. SunBrowser and FlowerBrowser were tested one after the other, each with two new profiles. For the two later Cloudflare tries, each engine used one profile.
• DICloak’s free plan allows 15 opens a day. Its 3 speed launches had no warm-up, and each one restored extra tabs, so its RAM figure includes those tabs.
• One speed launch of AliasMode Firefox did not load the probe page (a fault in our harness), and one launch of Donut Browser restored a tab from an earlier session. Each was repeated once.
• The Camoufox launch time includes the library’s own location lookup through the proxy, about 3 seconds.
• AliasMode Chromium ran from the AliasMode app source with the Linux engine build that ships in beta.70, before the release, on 2026-10-29. Its profiles started through the app’s Local API, which uses the same launcher as the app window.
• An early AliasMode Chromium run was discarded: another script on the host reached its control port. Later runs used a port that the script cannot reach. No other profile was touched.

We rerun a product when its vendor asks for a run on a paid plan through our contact page, or when our harness caused a failure. A product that cannot finish a check gets 0 for that check, with a note. Every rerun uses this method and gets a dated line in the changelog on the tier list page.

• We make AliasMode, one of the tested products.
• The list of impossible values comes from the checks that AliasMode runs before its own releases.
• The Cloudflare rule changed after the results.
• One computer with no GPU, one IP per product, and free plans or trials only. A Windows PC with a graphics card, or a paid plan, can give other results.