One fine morning Claude greeted me, like many others, with a message: “your account has been suspended.” At first I thought it was a glitch. Restarted. Not a glitch.
I was paying $200 a month, the most expensive plan. Kazakhstan citizenship, an account on a Kazakh phone number. None of it helped.
I contacted support the same day through the standard form. A couple of days later the answer came: we’ve checked everything, the decision is final. It reads as a polite “get lost”, a standard bot brush-off. In the same email they kindly offered to export all my data. I got excited, opened the archive — and found my email and phone number. That’s what “export all your data” means to them.
Along with the account went chats, projects, history, scheduled tasks, client tasks and analytics. Nothing can be restored, you can only start over.
This article is about what I was doing wrong. If you have one account with one AI and all your work depends on it, you’re exactly where I was on October 1.
What broke overnight
I have a content pipeline. A bot takes in my thoughts, an AI does the research, proofreads my articles so the spelling holds up, and nudges me to record videos. Then a gateway publishes on schedule to Telegram, Threads, Instagram and the blog, and metrics are collected automatically.
That day only generation stopped. The bot, the idea queue, the schedule, publishing and metrics kept working as if nothing had happened. The system turned out to be built right: it didn’t depend on whether a particular model was alive. The only part that died was the one where I had trusted someone else’s account to do the thinking. 😶😶😶
Why accounts get banned — and why it’s rarely about you
The official reasons: a usage policy violation, an account from an unsupported region, a breach of the user agreement. My email had a nice phrasing: “an internal investigation of suspicious signals associated with your account indicates a violation of our Usage Policy.” Which signals, which investigation — not a word.
In October 2026 there was a wave of bans for users from Russia and Hong Kong who were connecting through a VPN. Those regions have no official access, so formally it’s all legitimate. The unpleasant discovery: Anthropic works out your real location from a combination of signals — payment method, device, behavior. An IP that jumps between countries gives you away instantly.
I was absolutely sure that with an account registered to a Kazakh SIM card and a Kazakh passport there would be no problems. As it turned out — nope.
The appeal. You log in with the banned account and press the review request button. Like any corporation, they promise an answer within ten days, but I got mine literally the same day. Same wording: the decision is final. According to reviews, almost nobody from unsupported regions gets reinstated.
Payments are non-refundable by default. The exceptions: Europe and the UK, where the 14-day rule applies, and upgrades from a personal plan to a team plan. We, as it turned out, are not a European country. Though everyone gets the picture.
The subscription and the API are two separate wallets (at least there’s that). A paid plan doesn’t give you API access, and they’re banned separately. I wouldn’t have moved to the API anyway: with my tasks, clients and agents, the family budget would be gone in a week. So all my work sat on a single subscription account.
About OpenAI and Grok, honestly: I haven’t seen high-profile cases, but people get banned there too. It would be logical to switch to ChatGPT, but the chance of a ban isn’t zero — which means more wasted effort and time, and at the worst possible moment you’re back in the same situation. Grok, by the way, charges $0.05 for every request that violates its rules, and API credits aren’t refunded.
If you’re already banned: a step-by-step plan
Forget about creating a new account with a different email — you’ll be detected and banned even faster. So let’s go step by step:
Don't delete the app from your computer
Your computer or VPS still has artifacts, history, docs and files that will come in handy.
Request a review
The button on the restriction screen plus a short, emotionless note. The chance is small, but it's free.
Check your API key
It's a separate wallet: the subscription is dead, but the API may still be alive. Then you have a bridge for the move.
Set up another AI and pull everything off your disk
The key step. Don't move things by hand. Give the new AI this prompt:
"Go through my computer and server. Find everything related to setting up my AI agents: configs, rules files (AGENTS.md, CLAUDE.md), prompts, plugins, skills, automation scripts, integration settings for external services. Collect everything you find in one folder and make an inventory: what you found, where it was, and what it seems to have been for."
This prompt saved me a week: the agent pulled out most of what I had accumulated in the service over the years. Recovery went from "starting life over" to "putting things back on the shelves".
Plug what you collected into the new AI
The second prompt: everything collected in step 4 needs to be put to work. Rules and prompts in place, dependencies installed, access and integrations with external services set up. The model gets your context and picks up where everything stopped.
Test the restored setup on real work
Run your automation on the new model and watch the first run yourself. Whatever didn't come back, fix it using the inventory from step 4.
What not to do: write to support ten times. It’s pointless. I wrote once, and here’s what I got: support can neither name the reason for the ban nor lift it, all requests go only through the appeal form, and repeated messages may go unanswered. Conversation over, thanks for understanding.
And don’t start from a blank page: almost all of your old work is still with you.
How I chose a replacement
The ban happened in the morning, by lunch I was comparing options, and by evening I was working in a new AI, having lost almost nothing. I crossed out American services right away: a second ban right after you’ve moved everything means another painful migration. I looked at the Chinese ones: GLM, Qwen, DeepSeek.
I compared them on four things. How much the model remembers — the most important thing for agent work. Whether you can run it yourself, whether it works well with agent tools, and how much it costs. And payment: from Kazakhstan everything has to go through without jumping through hoops.
DeepSeek was out immediately: smarter people told me it hallucinates more than anything else on the market. Fine for a dumpling recipe, a disaster for serious projects.
I settled on GLM, and let me say the main thing up front: I build a lot, code is my main use case, and I chose the model primarily for how well it handles development. GLM holds the level of top American models on code, and when Anthropic itself got worried that someone else’s model had almost caught up with their flagship Mythos, that was the best signal for me: when a competitor acknowledges your replacement, you take it. Then the small stuff: payment from Kazakhstan goes through without issues, and ZCode, the app for Mac and Windows, looks almost exactly like the Claude app I was used to. Sat down and got to work.
Nobody guarantees the Chinese won’t ban you either. Any AI can decide it doesn’t like you anymore. But Chinese models haven’t restricted Russia in their rules.
A backup takes one evening: a step-by-step plan
Okay, but it’s smarter not to solve the problem when it hits, but to be ready in advance. So let’s prepare:
Set up automatic export
Every service has a data export. Put it on a monthly cron so the archive lands on your disk by itself. It doesn't save everything: chats can be exported, scheduled tasks and integrations with other services can't. After my ban I was offered an "export" of my email and phone number, so don't count on the service's export.
Build your own context folder
Rules, voice, projects and processes should live on your disk, not in the service's memory. For me it's a project folder with voice documents, prompts, platform rules and scripts.
Create a rules file
The open format is called AGENTS.md: a text file in the project root that describes how work is done in the project, what not to do and where things are. Tens of thousands of projects and agents read it. Claude Code has its own CLAUDE.md, looks for the nearest one up the folder tree and falls back to AGENTS.md if it has none.
Fold everything into one summary document
When I connect a new AI, I don't retell the project. I give it the folder and a file that briefly says what the system is, why it exists, how it's built, what's already decided and what to do next. The model reads it in a couple of minutes and works as if it had been there from the start.
Get a second vendor
A working account with another service, even on a free plan. You don't need to use it every day, but on ban day it should already be set up.
Install a local model
An AI that runs on your computer without internet and without an account can't be taken away. Heavy tasks are beyond it, but as an emergency option it saves the day. It installs via Ollama in half an hour.
Keep everything in one storage
The main rule I took away from the ban. All work results, all decisions, all knowledge should land in one place: my disk and my server. md files are updated as the work goes, nothing lives only in a chat. Then a new model gets the full context in a minute.
Evening's done, you're protected if:
- Export lands on your disk by itself, once a month, without you
- Rules, voice and prompts live in files next to the project
- The project has an AGENTS.md and a summary doc "what this is and how it works"
- A second vendor is set up and logged in, even if it's gathering dust
- A local model runs without internet
- No work result lives only in a chat
My next step, the one I’m eyeing: my own machine that stores and indexes the whole project context, with a local model managing that knowledge base. Reasoning and heavy lifting stay with cloud models — any of them. But the project memory stays at home.
A separate note on service memory. ChatGPT, Claude and Grok have their own memory features, but they’re vendor-locked: they don’t move between computers, and you can only take them out via export.
What to pay for and what not to
I didn’t consider the API for personal tasks: I’m not ready to pay for tokens every day. I need large limits, a large context, strong coding and work with subagents. A subscription fits that.
Prices as of early October 2026. Claude: Pro for $20, Max for $100 and $200 a month. ChatGPT: Go for $8, Plus for $20, Pro for $100 or $200. Grok has free access and a paid plan, but I can’t name exact prices — I didn’t find a clear page with numbers.
The GLM pricing looks like this: Lite with 10,000 credits a week for $18 a month, Pro with six times the limit for $80, and Max with fourteen times for $168. Quarterly billing gives 20% off, yearly 30% off: Lite comes to $12.60, Pro $56, Max $117.60. Note: even the top Max plan is cheaper than Claude’s most expensive plan, and Lite is cheaper than the $20 Pro. A nice bonus: from 13:00 to 9:00 Moscow time ZCode has a discount — tokens burn slower in that window, so night work costs almost half.
From personal experience, I hit daily and weekly limits at the same pace as with Claude (sometimes even slower), so in practice you won’t notice a difference.
The platforms compared in one table
I put what matters when choosing into one table. Monthly prices, without yearly discounts.
| Claude | ChatGPT | GLM | DeepSeek | Qwen | |
|---|---|---|---|---|---|
| Subscription | $20–200 | $8–200 | $18–168 | no | no |
| Free API | no | no | yes + 100M tokens a day | no | no, but open models |
| Context | 200K | 200K+ | 128–200K | 1M | 1M |
| Mac/Win app | yes | yes | yes, ZCode | yes | yes |
| Coding and agents | top | top | close to top | hallucinates | solid |
| Open license | no | no | no | MIT | Apache 2.0 |
| Bans users from the CIS | widely | rarely | haven’t heard | haven’t heard | haven’t heard |
Why I’m not buying my own hardware yet
At first I leaned toward a local model: your own box, nobody can ban anything. What caught my eye most was the NVIDIA DGX Spark: small, silent, 128 GB of unified memory.
But the price — around $7,000–8,000 — really bites.
And that wouldn’t be a problem if not for the second one. The box only handles models up to about 100 billion parameters; top models don’t fit even quantized. Meanwhile models only get heavier and hungrier. The “buy now, pay off in a couple of years” logic doesn’t work: in a couple of years new models won’t run on this box anyway.
A GPU setup of your own starts at about $12,000. Renting GPUs in the cloud costs $3,000 to $10,000 a month, and nobody refunds the idle hours.
The conclusion is simple: for now, stay on subscriptions and keep a local model as an emergency reserve.
Bonus: a free API the Americans don’t have
This is what struck me most in the whole story. Claude, ChatGPT and Grok have no free API at all: you want programmatic access, you pay for every token. GLM, on the other hand, has a free API with permanent models.
How I use it. Trivial tasks where it’s a pity to spend paid-plan tokens. Tasks where speed doesn’t matter: summaries, drafts, translations. And most importantly: when I burn through all the limits of the paid subscription, I switch to the slow free API and work doesn’t stop. With Claude at that point I just sat and waited for the window to reset.
Plus Z.ai throws in a bonus of 100 million tokens a day for the Flash model, though the freebie is only available for the first 5 days after paying for a subscription. The “ate all my tokens, now I wait” situation doesn’t happen at all. I press the Claim button in ZCode every morning:
Which models are free there: GLM-4.6-Flash and GLM-4.7-Flash. Performance-wise that’s roughly Claude Haiku 4 level, a year-old model with 128K context. It isn’t dumb and doesn’t hallucinate much, it’s just a bit weaker than the top ones. But again: it’s free.
I’m not a hardcore techie and didn’t run synthetic benchmarks — I don’t need them. I compared GLM on my everyday tasks: I tinker with code, design and development, and it handles them exactly the way Claude did.
A separate word on the interface. I’m used to working not in a terminal but in a Mac app. Something ultra-custom would have scared me off, but GLM has ZCode: almost the same as the Claude app.
Bottom line: a paid coding plan for the main work plus the free API as insurance and a cushion. The “GLM-5.3 in ZCode plus your own AGENTS.md” combo covers all my tasks for $168 instead of $200.