Mobile and 4G proxies for scraping: what a 400-request benchmark actually showed
A like-for-like test across Google Search, Amazon product pages, Instagram public profiles, and general websites sampled from the Tranco Top 1000 separates raw mobile IP access from managed scraping. The headline split is stark: conventional mobile networks handled most non-Google targets very well, while none returned a usable Google result page. A managed service using retries and, in its stealth mode, browser rendering and extra anti-bot handling, was the only configuration to retrieve Google results at scale — with much higher latency.
How the test was run
Six conventional mobile networks — Oxylabs, Decodo, SOAX, NodeMaven, DataImpulse and AirProxy — each received 400 requests: 100 per target group, in two rounds of the same 200-target set, sent sequentially with no client-side retries. The managed alternative was tested in Premium mode with retries, and separately with retries disabled, plus a stealth configuration run once for 200 requests because of higher credit cost.
The benchmark was written in Python using curl_cffi with Chrome impersonation. Returned HTML was parsed to check for expected content rather than trusting the HTTP status code. Raw results were saved after every request, together with failure details and response bodies when something went wrong. Once the 200 targets were ready, the list was frozen and every provider received the same URLs in the same deterministic shuffled order.
The general-web candidates were pre-checked before testing. Only pages returning real HTML, with a title and a reasonable amount of visible text, and not already serving an obvious CAPTCHA or anti-bot challenge were kept, so a site that was already broken would not be blamed on the proxy. The full Python benchmark, frozen target lists, raw results and result-processing scripts were made available in a GitHub repository, allowing the setup to be inspected, reproduced, or extended with another provider.
Success required usable content, not merely a 200 status. Google needed an actual results page; Amazon needed real product content; Instagram needed matching public-profile metadata; general sites needed a normal title and substantial visible text. Failures were split into blocked (target returned a challenge, empty or unusable page), error (timeout or technical failure), and provider-restricted (the proxy refused the connection before the target was reached). Speed is reported for successful requests only, as median and P95.
What the overall results mean
Google was the outlier. Every conventional mobile provider scored 0% for usable search results, whether rotating on every request or, for NodeMaven, switching from HTTP to SOCKS5. Premium managed retries reached 20% on Google; stealth reached 82%.
Outside Google, results were far stronger. AirProxy completed all 300 non-Google requests. NodeMaven and DataImpulse each completed 298 of 300. Oxylabs and Decodo were the fastest overall, with medians near 1.1 seconds, but provider-side refusals lowered their general-web scores. SOAX was slightly slower, with a 2.08-second median and 8.85-second P95.
A useful decision checklist follows from that pattern:
- Need raw mobile IPs, low latency and direct control for ordinary targets: prioritise AirProxy, NodeMaven or DataImpulse for reliability, or Oxylabs and Decodo for speed, after checking which domains each allows.
- Need a usable page from a heavily protected target: prioritise managed retries or stealth handling, accepting slower responses.
- Need Google search results specifically: do not rely on a fresh mobile IP alone; the benchmark found that approach returned no usable results.
- Need a new IP for every request on a dedicated SIM: budget for rotation time separately from page-fetch time.
Provider-by-provider notes
Oxylabs offers targeting by country, state, city, coordinates and ASN, with automatic rotation or sticky sessions. It succeeded on all Amazon and Instagram requests and was especially fast on Instagram, with a 0.78-second median. Its general-web raw score was reduced by 22 provider-restricted requests covering 11 domains in both rounds; on permitted general-web targets it succeeded 76 of 78 times. Google returned no usable results, with all 100 requests returning HTTP 200 without an actual SERP.
Decodo offers country, city and ASN targeting over HTTP(S) or SOCKS5, with rotating or sticky sessions. It also completed all Amazon and Instagram requests, with a 0.81-second Instagram median. It refused 22 general-web requests at proxy level; on permitted targets it succeeded 76 of 78 times. Google again returned no usable results, including one /sorry/ page.
SOAX offers country, region, city and ASN targeting with per-request rotation or sticky sessions. It completed all Amazon and Instagram requests and 90 of 100 general-web requests. Eight general-web failures were provider-restricted and two timed out; on permitted general-web targets it succeeded 90 of 92 times. Its overall median was 2.08 seconds and P95 was 8.85 seconds, with Amazon P95 at 11.64 seconds. Google returned no usable results: 56 requests landed on the sorry page and 44 returned HTTP 200 without search results.
NodeMaven combines residential and mobile access, with rotating or sticky sessions over HTTP(S) or SOCKS5. It completed all Amazon requests and 99 each on Instagram and general web, with no provider-side restrictions; its two non-Google failures were timeouts. Amazon median was 2.80 seconds with P95 14.36 seconds, while Instagram median was 0.90 seconds. A separate SOCKS5 rerun on Google also produced zero usable results in 100 requests, so protocol choice did not change that outcome.
DataImpulse offers country targeting with rotating or sticky sessions over HTTP(S) or SOCKS5. It completed all Amazon and Instagram requests and 98 general-web requests. The two general-web failures were the same provider-restricted domain, usda.gov; on permitted general-web targets it succeeded 98 of 98 times. Its median was 1.84 seconds and P95 was 4.98 seconds. Google returned no usable results: 88 requests returned HTTP 200 without a usable SERP, 11 landed on the sorry page, and one timed out.
AirProxy is Italy-only and uses dedicated SIM-backed 4G proxies rather than a shared global pool, with unlimited traffic and IP changes by schedule, manually or via API. Once a new IP was ready, it completed all 300 non-Google requests, with a 1.69-second median and 4.43-second P95. General web median was 1.16 seconds and Instagram median was 1.49 seconds. However, forcing a fresh IP before every request added roughly 10–20 seconds per rotation in this setup, and that overhead is excluded from the latency figures. Google returned no usable results in 100 requests.
Managed retries and stealth: reliability bought with latency
The managed Premium configuration used Italian geotargeting without JavaScript rendering, plus a Google-specific option for search requests. With retries enabled it completed 298 of 300 non-Google requests: Google 20 of 100, Amazon 98 of 100, Instagram 100 of 100, and general web 100 of 100. Google median was 8.32 seconds and P95 17.08 seconds; Amazon median was 6.35 seconds and P95 123.38 seconds; Instagram median was 1.33 seconds; general web median was 1.47 seconds. Overall there were 79 blocked and 3 errors in 400 requests.
Disabling retries dropped Amazon from 98% to 67%, Instagram from 100% to 92%, Google from 20% to 0%, and general web from 100% to 99%, with overall success falling from 79.5% to 64.5%. Errors rose from 3 to 42. Retries therefore rescued many first-attempt failures, but created a slow tail.
Stealth added JavaScript rendering and extra anti-bot handling, run for 200 requests. It reached 82% on Google (41 of 50), 98% on Amazon (49 of 50), 100% on Instagram (50 of 50) and 96% on general web (48 of 50), with 11 blocked and 1 error overall. Google median was 33.74 seconds and P95 80.76 seconds; Amazon median was 13.24 seconds; Instagram median was 11.72 seconds; general web median was 11.37 seconds. The trade-off is clear: managed handling improves the chance of receiving usable content on difficult targets, while raw proxy access is faster and gives the caller more control.
Pricing context and final choice
Entry prices captured at the time of testing varied widely and should be treated as dated, unverified reference points rather than current offers: Decodo and DataImpulse were the cheapest to start for a small test, NodeMaven and SOAX required modest prepaid or trial spend, Oxylabs' smallest mobile plan was higher, and AirProxy charged per dedicated proxy with a short trial option. The managed service used a credit model in which basic, Premium and Stealth requests cost progressively more credits.
Before choosing, check three things that headline success rates can hide: whether the provider allows your specific target domains, whether you need a fresh IP per request or can reuse one, and whether your bottleneck is IP reputation or anti-bot content delivery. Those distinctions explain most of the differences in this benchmark.
| Provider | Amazon | General web | Overall | Non-Google | Median | P95 | ||
|---|---|---|---|---|---|---|---|---|
| Managed Premium | 20.0% | 98.0% | 100.0% | 100.0% | 79.5% | 99.3% | 1.88s | 118.23s |
| Managed Stealth | 82.0% | 98.0% | 100.0% | 96.0% | 94.0% | 98.0% | 13.29s | 68.56s |
| AirProxy | 0.0% | 100.0% | 100.0% | 100.0% | 75.0% | 100.0% | 1.69s | 4.43s |
| NodeMaven | 0.0% | 100.0% | 99.0% | 99.0% | 74.5% | 99.3% | 1.82s | 12.46s |
| DataImpulse | 0.0% | 100.0% | 100.0% | 98.0% | 74.5% | 99.3% | 1.84s | 4.98s |
| SOAX | 0.0% | 100.0% | 100.0% | 90.0% | 72.5% | 96.7% | 2.08s | 8.85s |
| Oxylabs | 0.0% | 100.0% | 100.0% | 76.0% | 69.0% | 92.0% | 1.08s | 2.94s |
| Decodo | 0.0% | 100.0% | 100.0% | 76.0% | 69.0% | 92.0% | 1.10s | 2.66s |
Conventional providers: 400 requests each. Managed stealth: 200 requests. AirProxy latency excludes roughly 10–20 seconds of SIM rotation per request.
| Target | Requests | Success | Blocked | Errors | Provider-restricted | Success rate | Permitted success | Median | P95 |
|---|---|---|---|---|---|---|---|---|---|
| 100 | 0 | 100 | 0 | 0 | 0.0% | 0.0% | — | — | |
| Amazon | 100 | 100 | 0 | 0 | 0 | 100.0% | 100.0% | 2.37s | 3.14s |
| 100 | 100 | 0 | 0 | 0 | 100.0% | 100.0% | 0.78s | 1.21s | |
| General web | 100 | 76 | 1 | 23 | 22 | 76.0% | 97.4% | 0.92s | 2.35s |
| Total | 400 | 276 | 101 | 23 | 22 | 69.0% | 73.0% | 1.08s | 2.94s |
Oxylabs, 400 requests.
| Target | Requests | Success | Blocked | Errors | Provider-restricted | Success rate | Permitted success | Median | P95 |
|---|---|---|---|---|---|---|---|---|---|
| 100 | 0 | 100 | 0 | 0 | 0.0% | 0.0% | — | — | |
| Amazon | 100 | 100 | 0 | 0 | 0 | 100.0% | 100.0% | 2.18s | 2.79s |
| 100 | 100 | 0 | 0 | 0 | 100.0% | 100.0% | 0.81s | 1.28s | |
| General web | 100 | 76 | 2 | 22 | 22 | 76.0% | 97.4% | 0.86s | 2.05s |
| Total | 400 | 276 | 102 | 22 | 22 | 69.0% | 73.0% | 1.10s | 2.66s |
Decodo, 400 requests.
| Target | Requests | Success | Blocked | Errors | Success rate | Median | P95 |
|---|---|---|---|---|---|---|---|
| 100 | 20 | 77 | 3 | 20.0% | 8.32s | 17.08s | |
| Amazon | 100 | 98 | 2 | 0 | 98.0% | 6.35s | 123.38s |
| 100 | 100 | 0 | 0 | 100.0% | 1.33s | 2.58s | |
| General web | 100 | 100 | 0 | 0 | 100.0% | 1.47s | 2.99s |
| Total | 400 | 318 | 79 | 3 | 79.5% | 1.88s | 118.23s |
Managed Premium with retries, 400 requests.
| Target | Success rate with retries | Success rate without retries |
|---|---|---|
| 20.0% | 0.0% | |
| Amazon | 98.0% | 67.0% |
| 100.0% | 92.0% | |
| General web | 100.0% | 99.0% |
| Overall | 79.5% | 64.5% |
Same 400-request benchmark with retries disabled via transparent status code.
| Target | Requests | Success | Blocked | Errors | Success rate | Median | P95 |
|---|---|---|---|---|---|---|---|
| 50 | 41 | 9 | 0 | 82.0% | 33.74s | 80.76s | |
| Amazon | 50 | 49 | 1 | 0 | 98.0% | 13.24s | 17.35s |
| 50 | 50 | 0 | 0 | 100.0% | 11.72s | 23.23s | |
| General web | 50 | 48 | 1 | 1 | 96.0% | 11.37s | 29.74s |
| Total | 200 | 188 | 11 | 1 | 94.0% | 13.29s | 68.56s |
Managed Stealth, 200 requests, one round.
| Target | Requests | Success | Blocked | Errors | Provider-restricted | Success rate | Success on permitted targets | Median | P95 |
|---|---|---|---|---|---|---|---|---|---|
| 100 | 0 | 100 | 0 | 0 | 0.0% | 0.0% | — | — | |
| Amazon | 100 | 100 | 0 | 0 | 0 | 100.0% | 100.0% | 2.37s | 3.14s |
| 100 | 100 | 0 | 0 | 0 | 100.0% | 100.0% | 0.78s | 1.21s | |
| General web | 100 | 76 | 1 | 23 | 22 | 76.0% | 97.4% | 0.92s | 2.35s |
| Total | 400 | 276 | 101 | 23 | 22 | 69.0% | 73.0% | 1.08s | 2.94s |
| Target | Requests | Success | Blocked | Errors | Provider-restricted | Success rate | Success on permitted targets | Median | P95 |
|---|---|---|---|---|---|---|---|---|---|
| 100 | 0 | 100 | 0 | 0 | 0.0% | 0.0% | — | — | |
| Amazon | 100 | 100 | 0 | 0 | 0 | 100.0% | 100.0% | 2.18s | 2.79s |
| 100 | 100 | 0 | 0 | 0 | 100.0% | 100.0% | 0.81s | 1.28s | |
| General web | 100 | 76 | 2 | 22 | 22 | 76.0% | 97.4% | 0.86s | 2.05s |
| Total | 400 | 276 | 102 | 22 | 22 | 69.0% | 73.0% | 1.10s | 2.66s |
| Target | Requests | Success | Blocked | Errors | Restricted | Success rate | Permitted success | Median | P95 |
|---|---|---|---|---|---|---|---|---|---|
| 100 | 0 | 100 | 0 | 0 | 0.0% | 0.0% | — | — | |
| Amazon | 100 | 100 | 0 | 0 | 0 | 100.0% | 100.0% | 2.95s | 11.64s |
| 100 | 100 | 0 | 0 | 0 | 100.0% | 100.0% | 1.32s | 4.02s | |
| General web | 100 | 90 | 0 | 10 | 8 | 90.0% | 97.8% | 1.99s | 4.84s |
| Total | 400 | 290 | 100 | 10 | 8 | 72.5% | 74.0% | 2.08s | 8.85s |
SOAX, 400 requests.
| Target | Requests | Success | Blocked | Errors | Success rate | Median | P95 |
|---|---|---|---|---|---|---|---|
| 100 | 0 | 100 | 0 | 0.0% | — | — | |
| Amazon | 100 | 100 | 0 | 0 | 100.0% | 2.80s | 14.36s |
| 100 | 99 | 0 | 1 | 99.0% | 0.90s | 2.48s | |
| General web | 100 | 99 | 0 | 1 | 99.0% | 1.76s | 12.71s |
| Total | 400 | 298 | 100 | 2 | 74.5% | 1.82s | 12.46s |
NodeMaven HTTP, 400 requests.
| Protocol | Google requests | Successful SERPs |
|---|---|---|
| HTTP | 100 | 0 |
| SOCKS5 | 100 | 0 |
NodeMaven Google rerun, 100 requests per protocol.
| Target | Requests | Success | Blocked | Errors | Restricted | Success rate | Permitted success | Median | P95 |
|---|---|---|---|---|---|---|---|---|---|
| 100 | 0 | 99 | 1 | 0 | 0.0% | 0.0% | — | — | |
| Amazon | 100 | 100 | 0 | 0 | 0 | 100.0% | 100.0% | 3.32s | 4.98s |
| 100 | 100 | 0 | 0 | 0 | 100.0% | 100.0% | 1.56s | 3.41s | |
| General web | 100 | 98 | 0 | 2 | 2 | 98.0% | 100.0% | 1.27s | 5.70s |
| Total | 400 | 298 | 99 | 3 | 2 | 74.5% | 74.9% | 1.84s | 4.98s |
DataImpulse, 400 requests.
| Target | Requests | Success | Blocked | Errors | Restricted | Success rate | Permitted success | Median | P95 |
|---|---|---|---|---|---|---|---|---|---|
| 100 | 0 | 100 | 0 | 0 | 0.0% | 0.0% | — | — | |
| Amazon | 100 | 100 | 0 | 0 | 0 | 100.0% | 100.0% | 3.50s | 4.57s |
| 100 | 100 | 0 | 0 | 0 | 100.0% | 100.0% | 1.49s | 2.59s | |
| General web | 100 | 100 | 0 | 0 | 0 | 100.0% | 100.0% | 1.16s | 2.59s |
| Total | 400 | 300 | 100 | 0 | 0 | 75.0% | 75.0% | 1.69s | 4.43s |
AirProxy, 400 requests; latency excludes SIM rotation overhead.