I Benchmarked Just My Socks on 3 Nodes. The Numbers Were Fine — My Test Method Wasn't
I spent an evening benchmarking my Just My Socks subscription across three nodes, and the most interesting thing I found wasn't in the results table. It was that my first three test runs were measuring the wrong thing entirely — and nothing in the output told me so.
Here's the data first, then the trap, because the trap is the part that will save you an evening.
The results (evening peak, Nov 15, 2026)
All three runs went through the proxy, verified by checking that the exit IP matched the node IP before trusting any number.
| Node | Latency (client URL-Test) | ChatGPT | Claude | Gemini |
|---|---|---|---|---|
| s1 (Shadowsocks) | 663ms | 200, unlocked | 302 to login page | 200, unlocked |
| s3 (VLESS) | 199ms | 200, unlocked | 302 to login page | 200, unlocked |
| s4 (VLESS) | 218ms | 200, unlocked | 302 to login page | failed twice (timeout) |
A few things worth saying plainly about this table:
- ChatGPT was unlocked on all three nodes. No drama there.
-
Claude returned 302 to the login page on every node — and that is not a failure. A 302 to
/loginis what an anonymous visitor gets. The failure mode to watch for is a hard 403, which is what the same test returned when I accidentally ran it on my direct connection. If Claude 403s you, your exit IP is flagged; if it 302s you to login, the IP is fine and you just need to sign in. - Gemini is the one that actually varies by node. It worked on s1 and s3 and failed twice on s4. Same evening, same machine, same method. If Gemini is part of your workflow, node choice matters for it in a way it doesn't for ChatGPT.
- Latency is a snapshot, not a verdict. s1 measured 226ms in one run and 663ms in the next, minutes apart. s3 was 212ms then 199ms. Anyone showing you a single ping screenshot as proof a service is fast or slow is showing you weather, not climate.
The full table, with method notes, lives on my guide's hands-on review page. I keep it updated with dated runs instead of rewriting history.
The trap: my first three runs measured my direct connection
Before the table above existed, I ran the benchmark script three times and got identical results on three different nodes. That should have been impossible — different exit IPs should at least change the latency numbers. It didn't, because the script wasn't going through the proxy at all.
The script is Python, and Python's urllib on Windows ignores the http_proxy / https_proxy environment variables. It reads the proxy from the Windows registry (the system proxy setting) instead. My client sets up a local proxy on 127.0.0.1:10809, I set the env vars in PowerShell, ran the script — and it went direct every time. Three runs, three identical outputs, zero indication anything was wrong. The giveaway was switching to a completely different node — even a non-JMS one — and getting byte-identical results.
Two things fixed it:
- Verify the exit IP first, then trust the test. One command before anything else:
curl.exe -x http://127.0.0.1:10809 -s https://ifconfig.me/ip
If the IP it returns isn't your node's IP, your test traffic isn't going through the node and every number after it is fiction.
-
Test through the proxy explicitly.
curl -xforces the proxy regardless of what the runtime thinks the system proxy is. I also fixed the script itself to honor the env vars — the fix is in the repo — but the exit-IP check is the habit that matters, because it catches the problem no matter which tool you use.
If you benchmark any proxy service on Windows with a Python script, do the exit-IP check first. The failure mode is silent: plausible-looking numbers, wrong network.
How to reproduce this
Pick a node in your client, then in PowerShell:
curl.exe -x http://127.0.0.1:10809 -s https://ifconfig.me/ip
curl.exe -x http://127.0.0.1:10809 -s -o NUL -w 'ChatGPT %{http_code}\n' https://chatgpt.com/cdn-cgi/trace
curl.exe -x http://127.0.0.1:10809 -s -o NUL -w 'Claude %{http_code}\n' https://claude.ai
curl.exe -x http://127.0.0.1:10809 -s --max-time 10 -o NUL -w 'Gemini %{http_code}\n' https://gemini.google.com
200/301/302 on ChatGPT means unlocked. On Claude, 302 to login means the IP is fine; 403 means it's flagged. On Gemini, a timeout means that node doesn't reach it right now — switch nodes and retest before concluding anything about the service. And run the latency test in your client separately; the script's built-in latency check hits the provider's own subdomains, which is not the same as node latency.
One honest gap in my table: there's no packet-loss column yet. Neither the script nor the client's latency test reports it, and I won't invent numbers for it. If I find a clean way to measure per-node loss, the table gets a new column.
Disclosure: I maintain the justmysocks-guide project linked above. Some links in my guides are affiliate links; they do not change the prices or any of the test methodology described here.
Top comments (0)