Five proxy lists live on GitHub and get pulled into scripts every day. On 9 September 2026 we tested all five with one method, from one machine. Then we tested them again half an hour later. The spread runs from 20.8 percent down to zero. The largest file in the set holds 3,024 addresses. It relayed less often than a file a fifth its size.
The numbers
For each list we fetched the raw file the repository serves and parsed every address. Then we took a random sample of up to 120. Each address was opened, spoken to in its own protocol, and asked to fetch a real page. The reply had to carry an exit address that was not ours. A port check would not do. It cannot tell a proxy from anything else that answers.
| Repository | Addresses in the file | First run | Thirty minutes later |
|---|---|---|---|
| monosans/proxy-list | 619 | 20.8% | 12.5% |
| iplocate/free-proxy-list | 646 to 718 | 18.3% | 16.7% |
| TheSpeedX/PROXY-List | 3,024 | 7.5% | 5.0% |
| proxifly/free-proxy-list | 2,385 | 2.5% | 4.2% |
| clarketm/proxy-list | 400 | 0.8% | 0.0% |
The same test covered the wider market that afternoon. The first pass found 163 of 961 sampled entries relaying, which is 17.0 percent. The second found 14.0 percent.
We measured twice, and the numbers moved
Half an hour after the first pass we ran the whole audit again. Same machine, same method, same sample size.
Eight lists were measured twice with a usable sample. The median moved 3.3 percentage points. The largest move was 11.7. One list outside this table fell from 25.0 percent to 13.3. Two others moved upward instead.
With a sample of 120 the ordinary error near a 17 percent rate is about 3.4 points. Swings of eight to twelve points are therefore real churn, not noise.
Read every number here with that in mind. None of them is a verdict. They are two photographs of a market that reshuffles inside an hour. That is the strongest argument on this page. A list publishing a bare count, with no date and no method, is telling you nothing.
Size is not quality, and the file sizes prove it
The two largest files in the set are also two of the three weakest. TheSpeedX publishes 3,024 addresses and relayed 7.5 percent. Proxifly publishes 2,385 and relayed 2.5 percent.
Meanwhile monosans publishes 619 addresses, a fifth of TheSpeedX's file. It relayed 20.8 percent in the first run and 12.5 in the second. Even at its lower figure it beats both big files.
Work it through in absolute terms and the big file is still ahead. 7.5 percent of 3,024 is more working addresses than 20.8 percent of 619. That is a real advantage if you have the patience to find them.
What it costs you is everything else. To find one working proxy in the Proxifly file you will open about forty. In the monosans file you will open about five. Say you are on a schedule, or paying for compute, or trying not to hammer a target. Then that ratio is the whole story.
The claim that started this
Proxifly's repository states that proxies are fetched and validated every five minutes. Its README carried a live count when we read it.
We audited that claim on 5 September 2026, in the same hour its own file was stamped. Of the 2,211 entries it published, 1,770 accepted a connection and 306 relayed a real request. That is 13.8 percent against an advertised 2,326 working.
Its HTTP section was the striking part. It made up 71 percent of the file and 3.3 percent of it worked.
We are not calling that dishonest, and the distinction matters. Whatever their validation does, it is not equivalent to fetching a page through the proxy. A checker that opens a connection and stops keeps four entries for every working one. We measured that gap directly: 80.1 percent of that file accepted a connection and 13.8 percent proxied.
What these lists are genuinely good at
The distribution model is better than a web page and we will say so plainly.
A file in a repository is versioned, carries an obvious licence and needs no browser. Two lines of code will fetch it. Several of them split by protocol and by country. That is a real service to a developer. It is why these lists get used despite the hit rate.
The gap is not the format. None of them publishes what share of the file works. So you cannot choose between them without doing what we just did.
What no list can fix
We run a free pool of our own and watch what happens to these addresses. That lets us describe the lifecycle rather than guess at it.
Across 691,179 free proxies we have watched die, the median life was 144.5 hours, about six days. 22.6 percent died inside the first hour. 17.0 percent were still answering a month later. A file that is not re-verified constantly is wrong within a day, whoever publishes it.
Reliability is worse than the alive figure suggests. The median live free proxy in our pool succeeds on 45.5 percent of its own checks. Only 401 of 9,452 live entries met a bar of ten checks at 90 percent success.
The lists also overlap far more than their branding suggests. 116 sources feed our pool. The average live address appears on 17.86 of them, and 94.7 percent appear on more than one. These are not competing inventories. They are the same pool, re-published.
The part that breaks scrapers
Of the live free proxies we hold fraud data for, 89.3 percent were flagged for recent abuse. The average fraud score was 72.5, and none of them were residential. 82.9 percent sat on datacenter address space.
That is why an address can pass every technical check you run and still fail. The target already knows it. It was burned before you found it, by someone else, doing something else. No list can fix that, because the problem is not the list.
Free lists are the right tool for a test, a one-off fetch or learning how rotation behaves. Our own free list is served as txt, JSON or CSV with no key, for exactly that. When a job has to keep running, HProxy paid proxies are the alternative. They use residential, ISP, mobile and datacenter addresses that are not already flagged. They are billed by the gigabyte with no subscription, from $0.44/GB.
Method and limits
Everything here is one measurement, taken twice on 9 September 2026 from one datacentre in Germany. The 5 September audit is marked where it appears. A home connection may see different results.
We sampled up to 120 entries per list rather than testing every file in full. Each figure therefore carries ordinary sampling error. The two runs show the churn sitting on top of it.
We tested the HTTP files where a repository splits by protocol. A repository's SOCKS files may behave differently and were not covered.
These files change every few minutes. That is the point of them. It is also why this page carries a date rather than a verdict. We will re-run the test and republish under a new date.
Sources
- Our own market audit of the public proxy lists, run twice on 9 September 2026, and the single-list audit of 5 September 2026. Both instruments and their raw output are kept with this article's research folder.
- Our own free proxy pool: 763,409 addresses seen since it opened, 691,179 with a recorded death, queried 5 September 2026.