
For Chinese developers, a Claude ban notice arrives like bad weather: sudden, unexplained, and seemingly random. On LINUX DO, one of China's busiest developer forums, "存活挑战" (survival challenge) threads — where users report how many days their Claude account lived before being banned — have become a recurring ritual. One developer got tired of the folklore and decided to replace superstition with numbers. He built a questionnaire, a data dashboard, an IP-risk checker, open-sourced the whole project on GitHub, and collected 42 valid responses.
The result is a small, rough, fascinating dataset: a community trying to reverse-engineer Anthropic's risk engine from the outside. Everything below that comes from the forum thread is attributed as such — it's one community's self-reported experience, not a controlled study. Where a claim is independently verifiable, we've checked it against primary sources.
The survey organizer's own summary estimate: even if you do everything right — clean environment, compliant usage, no tricks — there is still roughly a 20% chance your account gets banned. That number is his subjective read of 42 responses, not a statistical result, and he says so. But it sets the tone for everything else: bans are treated as a cost of doing business, not an avoidable error. Several survey respondents reported losing paid accounts, which is why the thread's practical takeaway isn't "how to avoid bans" but "how to lose less when they happen."
Take the 20% figure as a mood, not a measurement. Still, it explains why a community that switched to Claude Pro in a visible wave now obsesses over ban survival mechanics.
The strongest-looking pattern in the data: the newer the account, the higher the ban probability. Older accounts not only survived longer — when banned, they were also more likely to win an appeal. (Only one successful appeal appeared in the sample, so treat the appeal claim as anecdotal.)
The folk wisdom that grew from this: register Claude accounts early, and prefer signing in through an old Google account rather than creating a fresh email login. This one got challenged in the comments, and the challenge is worth repeating: one member pointed out that Google's OAuth flow doesn't pass registration-date information to Anthropic at all, so "old Google account" is likely superstition rather than mechanism. It's a nice case of the community debugging its own myths in real time.
Reported-speech, but interesting: the account-type data points were brutal at the extremes. Purchased ready-made accounts were banned 2 out of 2; accounts created via Apple login were banned 3 out of 3. With samples that small these are conversation starters, not conclusions — but they rhyme with what Anthropic's terms actually say (more on that below).
Three popular beliefs failed to survive contact with the data:
Paying doesn't protect you. Across payment methods, ban probability hovered around the same baseline. Spending more — upgrading from Pro to Max — buys usage, not immunity. (For reference, Anthropic's consumer line-up runs Free, Pro at $20/month, and Max at $100/month for 5x Pro usage or $200/month for 20x, per multiple 2026 pricing trackers. "20X" in the thread refers to the $200 Max 20x tier.)
Static IPs didn't save anyone. Counter to community lore, having a static IP didn't significantly lower ban probability. The self-built VPS group actually showed the best survival numbers. What mattered more was hygiene: whether the IP was flagged by risk databases, and whether there was any leak — DNS leaks in particular correlated with very high ban rates. The organizer's rule of thumb: if a new account survives its first week, your network environment is probably fine. In the survey's terms, accounts registered in "problematic environments" were banned within seven days about 70% of the time, versus about 50% in clean ones. Small sample, big caveat — but the 7-day survival heuristic is the kind of actionable folk metric the community will actually use.
Device fingerprints were mostly noise. Most fingerprint items sat within about 10% of the baseline ban rate. The exception, again, was leaks — signs that the environment wasn't what it claimed to be.
The organizer's conclusion: if an account survives seven days, the remaining risk is almost entirely about usage. And the high-risk behaviors his data pointed at map neatly onto Anthropic's actual published rules:
The most intriguing reported finding wasn't in the survey at all — it was a comment. One developer reported that his team opened 15 identical Max 20x accounts in the same environment and with the same usage patterns. Seven were used for tuning Huawei Ascend AI accelerator cards and writing Ascend operators; eight were used for NVIDIA-ecosystem development. By the holiday week, all seven Ascend accounts were banned, while only one of the eight NVIDIA accounts was. He reads this as evidence that Anthropic reviews conversation content and applies risk scoring to specific topics or industries. The organizer agreed it fits the "usage patterns" dimension. It's one anecdote — but it's the kind of anecdote that makes you think about what large-scale AI safety filtering actually looks like from the receiving end.
Then there's the paranoia corner. One commenter warned against using Claude on a Mac over Wi-Fi, noting that on macOS the list of nearby Wi-Fi networks is easy to query without permission — if Claude's client can see SSIDs starting with CMCC or ChinaNet, your location is effectively disclosed. That's speculation, stated as such by the author, but it captures the ambient anxiety of the whole discussion: every signal might be a signal.
Verified against primary sources (Anthropic's Consumer Terms of Service and Acceptable Use Policy), the enforceable rules relevant to the thread:
Nothing in the terms promises users an explanation when enforcement happens — which is exactly why communities build folk theories in the first place. One banned user recently described the experience as "guilty as charged, charges unknown," clicking through to a policy page with no indication of which rule was broken.
Since the consensus is that bans can't be fully avoided, the thread's practical section is about damage control, all reported-speech:
The survey project itself deserves a mention as engineering: questionnaire, dashboard, an IP-risk checker built on multiple cross-referenced databases, the whole thing deployed on Cloudflare Workers with version tags and a public privacy policy. The folk science is scrappy; the tooling isn't.
A 42-person survey can't reverse-engineer a black-box risk engine. The organizer knows it; the thread knows it. But read as culture rather than statistics, the thread is revealing: a developer community treating platform risk as a collective engineering problem, building instruments, publishing data, and arguing with each other about which correlations survive scrutiny. The myth about old Google accounts got debunked by its own audience within an hour.
That's the healthiest possible version of ban paranoia: hypotheses with numbers attached, held lightly, updated in public. And if the 20% figure is even roughly right, it tells you something about the economics of AI subscriptions in China — which is why some developers have turned to third-party LLM relay stations as an alternative to keeping their own accounts alive. Sometimes the most rational response to an unexplained ban rate isn't better opsec. It's a different business model.