The Claude Anti-Ban Guide That Account Resellers Want to Buy for 50,000 Yuan — V2.0

Last week I released version 1.0, which received a lot of positive feedback and many comments and suggestions. Today I’m releasing version 2.0. So far, none of my accounts have been banned.

Let me start with a counterintuitive conclusion: Anthropic’s current approach is to assess account risk through a comprehensive judgment system. Their focus has shifted from preventing Chinese users to preventing abuse, exploiting the system for freebies, and violating their “values.” In other words, being Chinese is apparently no longer as big a threat as it used to be.

Accounts that are hard to get banned are ones that have been long-established, lightly used, valuable, and low-risk. Let me elaborate on each point.

(1) About Switching Languages, Home Broadband, Clean IPs, and DNS

I think these measures do have some use, but their weight in the risk assessment is much lower now.

Here’s what I mean: as long as your VPN connects to servers run by well-known, legitimate overseas providers, using Claude on such servers is a perfectly normal behavior. After all, plenty of foreigners do programming or ops-related work on servers too.

Also, don’t assume that changing your time zone and preventing DNS leaks means they can’t figure out who you are. There’s probably an AI agent behind the scenes that analyzes all kinds of signals — your coding habits, how you phrase questions, all sorts of traces of your company embedded in your code. So no matter what you change, it’s very hard to escape detection, unless your habits closely resemble those of a foreigner, and your server is genuinely located overseas and only interacts with overseas machines using their services. But that would be extremely inconvenient.

My accounts have always had the time zone set to Shanghai, and I use them on a Mac mini and a Mac, with both Chinese and English operating systems. Sometimes I even forget to turn on my VPN, and the Claude client will just return a 403 saying my country is not supported. However, this hasn’t caused any major impact — my accounts still haven’t been banned.

That said, buying home broadband or some reasonably good 🪜 still seems worthwhile. A relatively clean IP can significantly reduce the chance of being flagged. An IP address that’s been used by tons of Chinese users, or even by various relay sites, will almost certainly get banned. There’s no need to be paranoid about this, and this shouldn’t be seen as the only factor that gets your account banned.

Here’s what my IP situation looks like.

(2) Account Subscription, Registration, and Account Cultivation

Here, Kazyke teacher and other well-known bloggers have already reached a conclusion: using Google login or Apple login, and subscribing through the Google Play Store or the App Store, results in a much lower probability of being banned.

The reason is that the subscription prices on Google and Apple app stores are relatively expensive (a $100 subscription costs about $125, and a $200 subscription costs about $250). Also, since these are major service providers, they have more robust safety nets, including anti-stolen-card mechanisms. Abusive relay sites and institutions are unlikely to register and subscribe through these channels, so Anthropic considers the overall risk from these channels to be lower.

For account registration, most of my unbanned accounts were created by using the Claude app on an Apple or Android phone and logging in with Google directly on the phone. I find this method relatively stable, and it’s less likely to result in a ban.

As for account cultivation, my conservative approach is to use a new account for free for about a month first, then upgrade to the $20 subscription, and only move up to the $100 MAX after 1–2 months. If you think this process is too tedious and you don’t want to bother nurturing your account, at least don’t immediately activate a $200 subscription when you register a new account. Instead, start with the free tier, use it simply, then upgrade to $20, and after using up the $20 plan, move up to $100. Gradually step up, creating the impression that you’re a regular user whose usage is significant but whose upgrade path goes from cheap to expensive — like a normal user’s journey.

Additionally, speaking for myself, I don’t recommend beginners going straight to the $200 tier either. Claude is genuinely the most cost-effective, cheapest model to subscribe to right now. I sleep only five or six hours a day, keep my computer on year-round, and I still can’t use up the $100 quota.

(3) About Account Value

Our leading theory is the “account value” hypothesis. Account value means whether the content of your account can be “useful” to Claude. If the underlying judgment mechanism classifies it as “yes,” the probability of being banned drops dramatically.

We analyzed the ban rates of different user groups again and found that regular programmers and content creators seem to be banned quite easily. However, researchers who frequently publish papers, or mathematicians who do math research, are rarely reported as being banned.

Additionally, many programmers have noted that the timing of bans is also inconsistent. For example, many new accounts like to wait until you’ve finished a complete project before banning you. Meanwhile, people who do self-media content or simple projects can easily get banned on the same day they open their account. So if you’re asking ordinary questions, they like to ban you. If you’re asking cutting-edge questions that could contribute to the model, you’ll get banned later — or not at all.

The main character (me) is a graduate student in industrial engineering. In one class, the professor brought back a US textbook and had us model and analyze real-world American industrial production problems.

Because it was quite difficult, I used a new Claude account to ask about approaches for this problem, and ultimately Claude didn’t get it right either. Another account was used to assist with calculations for my modeling formulas during homework — also fairly cutting-edge topics. I suspect that around that time, Claude flagged me as some kind of industrial researcher or graduate/PhD student.

After graduating, I built ghfind.com. Although on the surface it grades programmers, the core methodology uses industrial engineering concepts and methods. This is also likely one reason why Claude’s upstream systems may have judged me as long-term valuable.

Therefore, it’s best to start right at the beginning when you open a new account — find some valuable, publicly available materials or papers to feed it. Note: don’t download papers from CNKI or any other paid channels — use freely accessible sources instead. Then have relatively in-depth conversations with the AI, and gradually upgrade to the $20 membership. This will likely significantly reduce the chance of being banned.

So, Anthropic is truly evil. The code you write with it, the papers you analyze with it, the homework you do with it — it takes all of that to train its next generation of models. That’s also why a large number of US government agencies, even the military, dare not cooperate with it right now. I think that’s partly the reason — it’s just too evil. Everyone should be sure not to use it for anything sensitive.

(4) About Usage Volume

The less an account is used, the harder it is to get banned. I feel this is also a confirmed fact. The main character has always been on the $100 plan, and for some reason I use it so frugally — possibly because I spend most of my time using domestic Chinese models, and only use Claude for open-source work and papers. My weekly usage stays below 60%, with only the occasional one or two weeks where I use it heavily, mostly when I wanted to try Fable. This also applies to OpenAI — if you often can’t finish using OpenAI’s models before a reset, it seems the probability of being “degraded” in routing is also greatly reduced.

I estimate that if you’re a new account that immediately activates the $200 tier and uses it all up right away, you’d very likely be flagged as a risky account and banned within a day or even less than 24 hours.

(5) Other Possibilities

There are also a few other tips and possibilities. The core principle is to not let Claude judge you as a risk account — even though some of its flags may be false positives, and its false-positive rate is quite high.

For example, Claude has recently been running promotional activities, like bonus quota giveaways. If your account has been registered for less than six months to a year, it’s best not to participate in these activities, as your account could easily be moved from the safe pool or the low-risk pool directly into the high-risk pool and get banned.

As for using the mobile version or the Claude client, these currently don’t seem to be strongly related to getting banned. In fact, I think using the desktop version might even be fine, since the desktop app has relatively broad permissions — perhaps using the desktop app and Claude Code to make calls is actually relatively low-risk. But again, the same principle applies: if you’re a new account and 80–90% of your usage is for writing code, it’s probably better not to install those apps right now. Wait until the account has been nurtured for a few more months before exploring Claude’s more advanced and interesting features.

Postscript

In short, Claude really wants to maintain its competitive advantage through anti-distillation, anti-freebie exploitation, and anti-abuse measures. That’s the current focus of their banning policies. Whether their approach is right or wrong, how we evaluate it subjectively, that’s one thing. But if you want to use their models freely right now, honestly, you can only follow the rules above — for regular users.

Finally, I want to say — I haven’t subscribed to Claude for quite a while myself, but really, Opus 5.5 is too impressive, too generational a leap. For ordinary coding tasks, domestic models are genuinely sufficient. If you don’t have to use Claude, don’t use it. The whole process of subscribing or nurturing accounts is just too complicated.

Unless you’re writing papers or working on open-source projects — competitive fields like these — I really wouldn’t want to use it either. Everyone should support DeepSeek and domestic Chinese models more. I hope domestic models can get better. Opus 5.5 really gives me a feeling of despair and a generational gap.