Skip to content

Benchmarks

A dated stealth comparison and local performance measurements for resource routing, session reuse, and concurrent fetching.

Comparison

WebSkrap values are an October 25, 2026 snapshot. Stock Playwright and CloakBrowser values came from CloakBrowser's upstream README, read on September 29, 2026. WebSkrap values came from a live report generated with WEBSKRAP_BROWSER_CHANNEL=chromium WEBSKRAP_LIVE_VIRTUAL_DISPLAY=1 WEBSKRAP_LIVE_GPU=mesa WEBSKRAP_CHROMIUM_SANDBOX=0 xvfb-run -a python scripts/live_stealth_report.py --no-open --report-only. Both suites ran WebSkrap 3.0.0 on its own CDP client with Chrome for Testing 156.0.8078.47 on ARM64 Linux, on a datacenter IP with no GPU and no proxy, rendering WebGL through Mesa lavapipe. The headed suite ran under xvfb-run; the headless suite used virtual_display=True, so both drove a headed browser on an Xvfb screen. The projects were not tested side by side. Detection sites and browser versions change, so these results do not predict a current score.

Feature comparison

WebSkrap vs other stacks

FeaturePlaywrightplaywright-stealthundetected-chromedriverCamoufoxCloakBrowserWebSkrap
reCAPTCHA v3 score0.10.3-0.50.3-0.70.7-0.90.9Pass headed (>=0.7 gate); scored in virtual display
Cloudflare TurnstileFailSometimesSometimesPassPassPass (non-interactive token)
Patch levelNoneJS injectionConfig patchesC++ (Firefox)C++ (Chromium)Own CDP client + native Chrome options
Survives Chrome updatesN/ABreaks oftenBreaks oftenYesYesYes (stock Chrome, no patched driver)
MaintainedYesStaleStaleUnstableActiveActive project tests
Browser engineChromiumChromiumChromeFirefoxChromiumChrome/Chromium
Playwright APINativeNativeNo (Selenium)NoNativeNo (own Python API)

Detection services

Stock Playwright vs CloakBrowser vs WebSkrap 3.0 headed on Xvfb

Detection serviceStock PlaywrightCloakBrowserWebSkrap headedNotes
reCAPTCHA v30.1 (bot)0.9 (human)PASSWebSkrap asserts score >=0.7 when Google's demo returns one
Cloudflare Turnstile (non-interactive)FAILPASSPASSProduction sitekey issued a real token without interaction, headed and on the virtual display
FingerprintJS bot detectionDETECTEDPASSPASSdemo.fingerprint.com/web-scraping returns demo data
BrowserScan bot detectionDETECTEDNORMAL (4/4)PASS0 abnormal checks, headed and on the virtual display
bot.incolumitas.com13 fails1 failPASSOnly tolerated network/spec false positives
deviceandbrowserinfo.com6 true flags0 true flagsPASSisBot: false
bot.sannysoft.comDETECTEDNot listedPASSNo rows flagged; WebGL rendered by Mesa lavapipe (gpu="mesa") on a GPU-less host
BrowserLeaks WebRTCNot listedNot listedPASSNo private ICE candidate IPs exposed
BrowserLeaks Client HintsNot listedNot listedPASSNo HeadlessChrome token
TLS / JA3 visibilityMismatchIdentical to ChromePASSTLS/JA3/JA4 surface is visible; no proxy mismatch without proxy
DNS leak standard testNot listedNot listedPASSResolver rows are public; optional proxy country/IP expectations supported

October 25 WebSkrap live summary: 25 passed, 2 skipped, 0 failed. Headed: 17 passed, 1 skipped. Virtual display: 8 passed, 1 skipped. Both skips were Are You Headless, which returned 502 Bad Gateway; the suites skip a site that answers with a 5xx status. reCAPTCHA v3 returned a score in both suites. The Turnstile demo uses a production sitekey, and the suite rejects the dummy token that a Cloudflare test sitekey returns.

Performance

Measured on 2026-10-25 with WebSkrap 3.0.0 against a local HTTP server with delayed assets. No external sites were contacted. Every session used the stealth setup: headless=True and virtual_display=True, so Chromium ran headed on a private 1920x1080 Xvfb screen. The host was ARM64, running Python 3.14.7 and Chrome for Testing 156.0.8078.47. Its Chromium sandbox was unavailable, so these runs used WEBSKRAP_CHROMIUM_SANDBOX=0 python benchmarks.py. Each figure is the median of three passes. Timings vary with hardware, browser version, and host load.

Resource routing

Full page load with delayed assets

PolicyTime (ms)vs ALL
DOCUMENTS174.400.53x
LITE204.320.62x
ALL328.441.00x

DOCUMENTS took 47% less time than ALL and LITE took 38% less. All three passes kept the same order.

Session reuse

Warm persistent session vs cold launch per fetch

ModeTime (ms)vs warm
Warm session reuse209.661.00x
Cold launch per fetch1,154.795.51x

Cold launch took 5.51 times as long as warm session reuse. Each cold fetch also starts its own Xvfb server, which adds to launch cost.

Concurrency

8 pages per batch from one session

160.22 ms

batch time divided by 8 pages

Median batch time was 1,281.78 ms (1,225.39 to 1,408.99 ms across passes). The per-page figure is a throughput equivalent, not the latency of an individual page.

Headed vs virtual display

Median time in ms, with the range across three passes

BenchmarkVirtual displayHeaded
ALL policy328.44301.09-351.45307.55304.13-309.37
LITE policy204.32199.26-233.81209.53195.46-211.47
DOCUMENTS policy174.40169.59-190.99169.05165.37-174.14
Warm session reuse209.66200.82-226.29208.64203.08-220.83
Cold launch per fetch1,154.791,137.44-1,216.781,021.96947.38-1,190.56
Concurrency, per page160.22153.17-176.12160.61157.73-167.13

Measured on 2026-10-25 on the same host, alternating three passes of python benchmarks.py with three passes of xvfb-run -a python benchmarks.py --mode headed, both with WEBSKRAP_CHROMIUM_SANDBOX=0. Headed mode used headless=False on a 1920x1080 Xvfb screen started by xvfb-run instead of one per session. Once the browser is running, the two modes differ by less than the spread between passes. Cold launch varied the most.

Each result averages 20 measured runs after two warm-up runs. See benchmarks.py for methodology.