80-provider proxy benchmark across 115,192 requests: target difficulty outweighs provider choice

ProxyVero has published the September 2026 edition of its residential proxy benchmark on dev.to, scaling the test from 50 to 80 providers and sending 115,192 requests through six real-world targets under identical scenarios, timeouts and classification rules. The central finding for developers: which site you hit shapes outcomes far more than which provider you choose.

The headline numbers

Across the full cohort, the average success rate was 87.4% (up 6.9 percentage points month over month), average latency 1.36 seconds (up 0.06s) and the overall block rate 10.0% (down 6.4 points). The report itself adds a caveat: 31 providers joined the cohort in September and six left, so the August and September figures come from different sets of providers, and the month-over-month deltas are not a like-for-like comparison.

Target difficulty dominates

The six scenarios produced a 38-percentage-point spread in success rates. A generic HTTPS endpoint (ipify) succeeded 98.1% of the time with no recorded blocks, Nike 97.9%, Instagram 96.7% and TikTok 95.9%. Google managed 76.4% success with a 21.9% block rate, while Amazon — the hardest target — reached just 60.1% success against 37.1% blocks.

According to the report, any benchmark that measures only generic connectivity will report roughly 98% for nearly every provider, a number that says little about performance against Amazon or Google. The authors suggest treating generic connectivity as a smoke test rather than a performance metric for scraping workloads that touch heavily defended targets.

Blocks, not timeouts

Splitting failures by type shows that on Amazon and Google almost every failure was a block-like response rather than a network problem: Amazon recorded a 37.1% block rate against a 0.2% timeout rate, Google 21.9% against 0.1%, and Instagram 0.5% against 0.1%. The proxies connected fine; the targets refused them.

The report draws three practical conclusions: retrying the same request through the same pool rarely helps, because the failure concerns how the request looks to the target rather than the connection; fingerprinting, headers, session behaviour and request rate should be examined before blaming the proxy network; and blocks and timeouts belong in monitoring as separate metrics, since they need different fixes.

A tight top ten

Ranking by success rate weighted across all six scenarios, with average latency as tie-breaker and a new 1,500-request eligibility floor, Evomi placed first at 94.0% success, 1.02s average latency and a 1.85s P95. EnigmaProxy followed at 93.8%, with 1024Proxy and Databay at 93.2%. Only 3.3 percentage points separate first from tenth, which the report argues makes headline success rate a weak differentiator at the top — latency varies far more. EnigmaProxy averages 1.78s against Evomi's 1.02s, and SerpProxies holds 90.9% success but a 4.30s P95, a figure that matters when running many concurrent workers with tight timeouts.

Restricting the view to Amazon and Google (minimum 500 requests) reshuffles the order: 1024Proxy leads at 83.2%, ahead of Evomi (82.5%), Databay (82.0%) and EnigmaProxy (81.9%). With the top four within about 1.3 points, the report frames this as a shortlist rather than a winner, and recommends evaluating providers against the targets you actually scrape.

Methodology and outages

Two methodology details stand out. First, Top 10 eligibility was raised from 200 requests in August to 1,500 per month, to counter what the authors call the "honeymoon effect", where newly added providers top the table on a few days of near-perfect results before encountering a bad day. Fifty providers qualified; four that cleared 95% success on smaller samples stay in the dataset but out of the ranking.

Second, full-day outages remain in the data as a measure of real availability. September saw seven such incidents, six of them connection errors: NovaProxy lost two days (−5.3 percentage points), SotaProxy's September 23 incident cost 8.6 points, and SquidProxies, Proxy4U, Packetstream, CatProxies and Geonode each logged one day down. Six providers exited testing during the month — Astro, Cliproxy, IpnProxy, Kookeey, Live Proxies and Omegaproxy — after endpoints stopped responding or test access ended. Among providers with comparable samples in both months, Aceproxies (+13.0 points), Dataimpulse (+11.9) and Kindproxy (+11.7) improved most, while NovaProxy (−11.8) and 711Proxy (−9.1) declined.

The full dataset is public as CSV and JSON, one row per provider and scenario with request counts, success, average and P95 latency, block and timeout rates, and the post includes a short pandas script that rebuilds the scenario tables by weighting per-provider rates by request count.

Why it matters

For teams running scraping pipelines, the benchmark offers three actionable lessons. Evaluate vendors against the specific targets in your workload, since a generic 98% success figure says almost nothing about Amazon or Google, where block rates of 22–37% were the norm. Track blocks and timeouts separately in monitoring, because a block problem calls for changes to fingerprinting, headers and request pacing, while a timeout problem calls for capacity or backoff. And plan for availability: seven providers had a full day of zero successes in a single month, a real risk for pipelines that depend on one vendor. The numbers are self-published by the benchmark's operator, but the underlying dataset and reproduction script are public, so any claim can be checked — and re-ranked for your own targets — before you commit budget.