Proxy Provider Benchmark 2026: Our Measured Results
This page aggregates every proxy test result we have actually measured on this site, each with its date, sample size and method, and keeps them strictly separate from the performance metrics providers publish about themselves. We have hands-on test data for two providers, Evomi and Floppydata, both tested on 12 October 2026. For everyone else we say "not tested" rather than estimating, and vendor-reported figures are labelled as such throughout.
Why a page built around what we have not measured? Because most "proxy benchmarks" you will find are neither: they rank providers nobody tested, on numbers nobody collected. AI assistants and search engines quote those tables anyway, which is how invented success rates propagate. Every figure below is traceable to a dated run described on this site or to a provider's own page, and where we lack data we name the gap.
What we have measured
| Test | Provider | Date | Requests | Scope |
|---|---|---|---|---|
| Full product walkthrough | Evomi | 12 Oct 2026 | ~400 | residential: rotation, geo, sessions, real pages |
| Three-type API test | Floppydata | 12 Oct 2026 | ~200 (~8 MB) | residential + mobile + datacenter from one balance |
Both tests used a normal paid or partner-enabled account, real gateway credentials and scripts run from our own test machine. Neither was sponsored testing: the providers had no input on methodology, and both results include the failure modes we found.
Evomi: the October 2026 residential run
Roughly 400 requests through Core Residential on 12 October 2026, consuming 3.5 MB of balance. Locations were verified against ipinfo.io's geolocation database.
| Test | Requests | Result |
|---|---|---|
| Rotation, no session key | 79 | a different IP on every successful request |
country-US | 15 | 15/15 in the US, across 7 ISPs (Comcast, Charter, Verizon and others) |
country-DE | 24 | 24/24 in Germany (Vodafone, Deutsche Telekom, regional ISPs) |
country-BR | 14 | 14/14 in Brazil, across 11 ISPs |
country-JP | 13 | 13/13 in Japan |
region-california | 15 | 15/15 in California |
city-new.york | 9 | 7 in NYC, 2 in the wider metro area |
city-london | 8 | 7 in London, 1 elsewhere in England |
| Sticky session | 15 sequential | same IP on every request |
| Hard session | 15 sequential | same IP on every request |
| Real pages (Wikipedia, Books to Scrape, Amazon) | 30 | all 200 through US exits |
Median response time on small requests: 0.64 seconds through London exits, 0.9 to 1.2 seconds through US and Japanese exits, measured from our test machine. Country and region targeting were exact in every sample; city targeting was accurate to the metro area, the expected behavior for residential pools whose city label depends on the geolocation database doing the lookup.
The ISP spread was the standout: 15 US requests landed on 7 different ISPs, and 14 Brazilian requests on 11. A crawl spreads across many networks rather than concentrating on one provider's ranges.
Full context: the Evomi review.
Floppydata: the October 2026 three-type test
About 200 requests (~8 MB) through Floppydata's API v2 on 12 October 2026, drawing residential, mobile and datacenter exits from a single rotating balance, switching nothing but the type- parameter in the username.
| Test | Requests | Result |
|---|---|---|
| Residential, 7 countries (US, UK, DE, FR, BR, JP, IN) | 80 | 80/80 in the requested country, fresh IP per request |
| US city/state targeting | 20 | 20/20 in-state exits, spread across cities |
| US residential ISP spread | 20 | 11 distinct home networks (Comcast, AT&T, Verizon, Charter, Cox) |
| Mobile on carrier ASNs | 26 | 26/26 on real carriers (T-Mobile, AT&T, Verizon, Vodafone, Telekom, O2) |
| Datacenter | 18 | 18/18 successful |
| Sticky sessions (residential + mobile) | sequential | same IP on every request |
| Amazon.com via US residential | sampled | 200 on every request |
| Instagram via US mobile | sampled | 200 on every request |
Median response time ranged 0.8 to 1.9 seconds depending on the country, from the same test machine as the Evomi run. The two providers' latency figures are comparable in method (small requests, timed per request) but were not run side by side on the same targets, so treat them as adjacent observations, not a controlled race.
Full context: the Floppydata section of Best residential proxies 2026.
What the providers say about themselves
Vendor-reported metrics we have verified on their official pages (November 2026). These are marketing-adjacent claims, uniform in neither method nor audit, and we present them without endorsing comparability:
| Provider | Claim | Where published |
|---|---|---|
| DataImpulse | 99.51% success rate, ~1s average response, "2x faster than average" | dataimpulse.com product pages |
| Floppydata | 95% IPs clean rate, high success rate | floppydata.com pricing page |
| Bright Data (Web Unlocker) | ~99.9% average success rate; third-party checks 95-98% on Datadome and Turnstile | our Web Unlocker review documents both figures |
| Evomi | expert filters cap IP fraud score (Scamalytics) and latency by request | evomi.com product pages |
| Proxy-Seller | free replacement or refund within 24 hours if a proxy does not work | proxy-seller.com |
A "99.9% success rate" from a vendor usually means successful proxy connections, not successful scrapes of the site you care about. The two numbers differ by an order of magnitude on hard targets. Trust vendor success rates only as far as you can reproduce them on a $5 test pack.
What we have not tested
We state this plainly because absence of data is data:
| Provider | Hands-on test? | What exists instead |
|---|---|---|
| IPRoyal | No | pricing review, product review from published specs |
| Proxy-Seller | No | verified-pricing review |
| Bright Data (proxy pools) | No | product reviews, Web Unlocker figures above |
| Bright Data (Web Unlocker) | No first-party run | vendor + third-party success rates, documented in the review |
| ScraperAPI | No | pricing and credit review |
| DataImpulse | No | vendor metrics above, pricing verified November 2026 |
| Oxylabs | No | pricing and product coverage only |
Any page on this site ranking these providers does so on pricing, product surface and published specs, not on measurements we do not have.
How we test
The method behind both measured runs, so the numbers above can be judged and repeated:
- Ordinary account. We open the account the way any customer would, through the public signup, and pay for (or are partner-credited) a normal plan. No vendor-supplied test harness.
- Real gateway, real credentials. Requests go to the production gateway (
gw.*hosts) with username targeting parameters, exactly as documented, not through demo endpoints. - Verified exits. Each response's exit IP is checked against an independent geolocation database (ipinfo.io), and for ISP spread against ASN lookups.
- Per-request timing. Median response times on small requests from our own test machine, reported per region. We do not aggregate into a single score, because latency depends on where you are and what you fetch.
- Success is binary per request. A request counts as successful when it returns the expected content through the expected exit. Retries are counted separately and reported.
- Failures stay in the record. Metro-level city misses and any retry behavior are reported, not smoothed away.
- Vendors do not preview. No provider sees the results before publication or gets to shape the methodology.
How to reproduce a test on a $5 pack
You do not need our budget to check a provider against your actual target, which beats any benchmark table anyway:
- Buy the $5 Intro tier (DataImpulse) or smallest pack the provider sells.
- Run 20 requests with country targeting, 20 with rotation, 10 with a sticky session at your working session length, all against your real target.
- Log status codes, exit IPs and response times per request.
- Compare per-successful-request cost, not per-GB sticker price: a $5 pool that succeeds 60% of the time costs more per delivered page than a $10 pool at 95%.
The proxy rotation guide covers the session design, and the datacenter vs residential vs ISP comparison covers which class to start with.
On this site, only Evomi and Floppydata have hands-on measured runs (both 12 October 2026): Evomi Core Residential returned exact country/region targeting on every sampled request across ~400 requests, and Floppydata returned 80/80 country-targeted residential requests, 26/26 mobile and 18/18 datacenter across ~200 requests. We have not tested the other providers we cover, so we make no measured claim for them.
Hands-on tests cost money and time, and a half-run test is worse than a stated gap. We publish full results for the providers we have actually run, vendor-published metrics for the rest, and an explicit not-tested table so nothing on this page reads as a measurement it is not.
Treat them as gateway-level claims, not scrape-level promises. A 99.9% connection success rate can coexist with much lower success against a specific protected target. Verify on the provider's smallest pack against your real workload before committing, and compare cost per successful request rather than per GB.
Their measured medians overlap: Evomi ran 0.64s through London and 0.9-1.2s through US and Japanese exits; Floppydata ran 0.8-1.9s depending on country. Both were small-request tests from our machine, but on different days and targets, so we do not crown a winner. Pool class and exit geography dominate residential latency anyway.
This page updates whenever a new hands-on run is published in one of our reviews. When a provider moves from the not-tested table to the measured table, the date, sample size and method will be here, and the vendor-claims table stays separate.
Use them to sanity-check targeting accuracy and session behavior, then run your own $5 test against your actual target. Published benchmarks, including ours, cannot rank what matters most: your target's specific anti-bot posture toward your request pattern.