If employees at your company are using Claude through VPNs or intermediaries, consider that access temporary: in early October 2026, Anthropic blocked another batch of accounts from unsupported regions, including long-running paid accounts from Russia. The company should not be looking for a new workaround; it should spend a week identifying where Claude is built into its workflows, exporting data, and preparing a backup model that can be used legally in Russia.
This material was prepared by AIrassvet using publicly available sources, verified on October 8, 2026. It is intended for business owners and executives, IT directors, and team leads in development, marketing, and support who already use Claude or Claude Code. This article is not legal advice: service terms and legal requirements should be checked against primary sources and with a lawyer. AIrassvet is not affiliated with Anthropic or any other companies mentioned, and the section at the end describes the publisher's services.
Contents
- What happened in October 2026
- Why this affects more than just individual users
- Why bypassing the block is not a business solution
- In short: what to do in the first seven days
- Step 1. Inventory: where you use Claude
- Step 2. Save data and context
- Step 3. Choose a backup setup
- Step 4. Portability: prompts, agents, code
- How to test the replacement before switching
- What the AI law changes
- Frequently asked questions
- Primary sources and verification method
What happened in October 2026
On October 1 and 2, 2026, users in Russia and Hong Kong began reporting widespread Claude account blocks. According to RB.ru, which cites the South China Morning Post, the restrictions also affected users who had been working for more than a year through a single foreign IP address and paying for a Max subscription at $200 per month. The official wording in the emails was "suspicious signals"; users are offered the option to appeal or export their chat history and delete the account.
What is known as of October 8, 2026:
- Russia, Hong Kong, mainland China, and Macau are not included in the list of regions where Claude is officially available. This is visible on the Anthropic Supported Regions page; Russia is not on the list.
- The blocks affect long-running accounts. The author of the Habr analysis It says the autumn wave has reached old personal accounts, and it reminds readers of the April case, when the agrotech company lost about 110 accounts overnight.
- It is not only the IP address that is checked. According to TAdviser, in addition to IP, the system analyzes additional parameters that make it possible to determine a user's location at the country level.
- There was no public announcement of a new policy. This was reported by both CNews and Habr. According to users, appeals are not always approved.
- This is not the first wave. CNews notes that in May 2026, Anthropic disabled several hundred profiles of Russian users without warning; at the time, people lost their chat history and code stored in the interface, and subscription fees were not refunded, according to Habr.
Amid the blocks, terms changed at other providers during the same week. As of October 9, Google will leave only the Flash-Lite model available to free users of Gemini apps; Flash and Pro are moving to paid plans (Habr, October 5, 2026). And Meta and Microsoft, according to a summary of The Information's report, are reducing employees' internal use of Claude in favor of their own tools (RSWebsols, October 6, 2026). The bottom line is the same: access to a specific foreign model is a vendor condition, not a company asset.
Why this affects more than just individual users
Claude is popular in Russian teams. According to RB.ru, citing foreign purchase data for summer 2026, Claude accounted for 5.3% of Russian transactions in foreign services, while ChatGPT accounted for 7.2%. Some of these subscriptions were set up by employees for work tasks, often without the IT department's knowledge.
For a company, a block is more than just one person's inconvenience:
- the process stops, if content drafting, customer replies, code reviews, or analytics depend on Claude;
- context is lost: projects, uploaded documents, saved instructions, and decision history;
- integrations break, if the API key or agent is embedded in a CRM, bot, or development pipeline;
- shadow IT appears: employees switch on their own to the first service they find and move work data there.
The main risk is not the block itself, but that the company does not know exactly where Claude is being used. So the plan starts with inventory, not with choosing a new model.
Why bypassing the block is not a business solution
Instructions are already appearing online for how to maintain access: your own server with a permanent foreign address, intermediaries, foreign cards. For a personal experiment, that is the user's choice, but you cannot build a company workflow on that for several reasons.
- Violation of the provider's terms. Using the service from an unsupported region violates Anthropic's rules; in this case, a block is an expected consequence, not a malfunction.
- Unpredictability. The author of the Habr analysis tested 25 services from a home Moscow IP on October 1, 2026: 9 worked, while the rest returned a regional restriction or did not respond. The author also warns that "any one of these doors can close." A process that depends on a door that can close on any day is not manageable.
- Personal and business data. If the prompts include customer or employee data, transferring it to a foreign service through an intermediary requires a separate legal assessment under Federal Law No. 152-FZ. For more details, see our article on rules for transferring customer data to AI.
- Opaque intermediary. In the tokenizer benchmark on Habr , the author found that one aggregator for two Claude models returned token counts matching OpenAI's tokenizer. The gateway may be recalculating tokens itself, or the requests may be going to the wrong model. There is no way to verify this from the inside.
In Brief: What to Do in the First Seven Days
| Day | Action | Result |
|---|---|---|
| 1 | Announce the rules to employees: do not create new accounts by bypassing controls, and do not upload work data to random services | Shadow AI growth stopped |
| 1–2 | Inventory: who is using Claude, where, and for what; what API keys and integrations exist | Dependency register |
| 2–3 | Export the history, projects, instructions, and templates from still-active accounts | Work context preserved |
| 3–5 | Select a backup model for each critical scenario and run a test with your own examples | Verified replacement |
| 5–6 | Move integrations to a layer where the model can be changed by configuration | Switching without rewriting code |
| 7 | Decide for each process: switch it, keep it with a backup, or move it into your own environment | Quarter plan |
This timeline is an editorial recommendation from AIrassvet; actual timing depends on the number of processes and integrations.
Step 1. Inventory: Where You Use Claude
Create a simple spreadsheet inventory. For each use case, record:
- process and owner: for example, “reviewing customer feedback — marketing,” “code review — engineering team”;
- access method: web interface, Claude Code, API key, intermediary, built-in feature in another product;
- what data is sent to the model: public text, trade secrets, personal data;
- criticality: what happens if access disappears tomorrow — inconvenience, delay, or process shutdown;
- what is stored only in the account: projects, instructions, decision history.
Also review code and configurations separately: API keys, environment variables, and SDK calls to a specific vendor. If the model is hard-coded in dozens of places in your code, switching will take weeks; if calls go through a single layer, it can take hours. About this approach — in the article on a unified LLM API for business.
Step 2. Save Data and Context
While the accounts are still working, export:
- dialogue history that is important for work — using the built-in data export in account settings;
- project instructions, system prompts, templates, and examples of successful responses;
- documents that were uploaded to projects, if their originals are not in the corporate storage;
- agent settings and rules for Claude Code, if they are stored outside the repository.
Move all of this into your corporate system: knowledge base, repository, shared drive. Instructions and examples are the main asset that moves between models; the model itself is replaceable, but the task description, quality criteria, and reference answers are not.
Step 3. Choose a Backup Environment
There are three replacement options, and different processes call for different ones.
| Option | When It Fits | Limitations |
|---|---|---|
| Russian models via API (GigaChat, YandexGPT, and models in Russian cloud services) | Text, support, document work in Russian, processes involving personal data | You need to test the quality on your own tasks; for code and complex agent chains, results may differ |
| Open models in your own environment | Sensitive data, security requirements, high stable request volume | GPU servers, maintenance, and updates are required |
| Foreign models through a legal and transparent channel | Tasks without personal data where the quality of a specific model is critical | The risk of changing terms still remains; you need a contract and a clear jurisdiction |
A few points from recent sources:
- Russian language and price. According to a measurement on Habr from October 1, 2026, YandexGPT and GigaChat use fewer tokens for Russian text than for an English translation, while in Claude Opus 5 almost every letter becomes a separate token. For Russian-language tasks, this directly affects the budget; see our detailed breakdown of why Russian text is more expensive in LLMs.
- Overview of Russian models — in the article LLMs in Russia in 2026: GigaChat, YandexGPT, Qwen, DeepSeek.
- Your own environment. A comparison of cloud costs and your own servers — in the article cloud or local LLM: total cost of ownership.
You do not have to choose one model for everything. A common setup is routing: simple, high-volume requests go to a cheaper model, complex ones go to a stronger model, and data with higher requirements stays only in your own environment. We described how to set this up in the article on an AI model router.
Step 4. Portability: prompts, agents, code
To keep the next vendor switch from turning into a crisis, establish three rules.
- Store prompts and instructions in a repository or knowledge basewith versioning and an owner. Not in the chat interface.
- Model calls through a single layer (a gateway or your own wrapper), where the model and vendor are set by configuration. Agent tools should be described in a neutral format, not tied to a single SDK.
- A set of benchmark tasks — 30 to 100 real examples with the expected result for each process. With that in place, any new model can be tested in a day.
For a detailed breakdown of the risks of vendor lock-in and how to prepare for shutdown, see our article exit plan from an AI service.
How to test a replacement before switching over
Before moving a process, run the benchmark tasks through the candidate models and compare:
| Criterion | How to measure |
|---|---|
| Quality | Share of answers that an expert accepts with no edits or only minor edits |
| Cost | Spend per typical request on your own texts, not the per-million-token price from the rate card |
| Speed | Response time for a typical request during business hours |
| Stability | Rate of errors and service failures during a one-week trial run |
| Data | Where requests are processed and stored, whether there is a contract, and the jurisdiction involved |
Methodology: editorial recommendation by AIrassvet. The decision is made by process, not for the whole company at once: one model may be suitable for customer replies, another for code, and for documents with personal data only a model in your own environment.
What the AI law changes
The Law on Supporting the Development of AI Technologies (243-FZ) introduces the concepts of sovereign and national large foundation models. According to a CNews report dated July 9, 2026, the cases where only sovereign and national models are allowed do not apply through September 1, 2032, to information systems with models created and operated before March 1, 2027.
For most commercial companies, this law does not impose direct bans on foreign models, but the trend is clear: for government systems and sensitive industries, requirements for model “Russianness” will keep increasing. If you work with public-sector clients or in a regulated industry, it is worth checking your backup setup against these requirements right away.
Frequently asked questions
Why is Claude blocking accounts from Russia?
Russia is not on the list of regions where Anthropic officially provides Claude. In October 2026, the company blocked a batch of accounts that were operating through VPNs or foreign servers, citing “suspicious signals.” Anthropic has not publicly announced a new policy.
Can a blocked Claude account be restored?
Anthropic says you can submit an appeal, but according to user reports, they are not always approved, and some decisions are described as final. For a company, it is safer to assume access will not be restored and export data from active accounts in advance.
Is it safe for a company to use Claude through a VPN or intermediary?
For business workflows, no: that kind of access violates the vendor’s rules, can disappear at any time, and transferring personal data to a foreign service through an intermediary requires a separate legal review under Federal Law 152-FZ. The smarter move is to migrate critical processes to a model that can be used legally.
What can businesses in Russia use instead of Claude?
It depends on the task: Russian models via API are suitable for Russian-language text and document workflows, while open models in your own environment are better for sensitive data and large volumes. The choice should be made for each process after running benchmark tasks.
How quickly can processes be moved from Claude to another model?
If model calls go through a single layer and instructions and benchmark examples are stored in a repository, switching takes hours or days. If the model is built directly into code and interfaces, the first step is to move calls into a shared layer.
Does this problem affect other foreign AI services?
Yes, for the same reasons. According to a Habr measurement from October 1, 2026, OpenAI, Anthropic, and Google responded to requests from a Russian IP with regional restrictions, and Google has been reducing free access to Gemini since October 9. Any foreign service can change its terms.
Primary sources and verification method
Checked on October 8, 2026:
- Anthropic, Supported Regions — the list of regions where Claude is available; Russia is not included.
- RB.ru — a summary of South China Morning Post coverage of the October 1, 2026 blocks and the share of foreign purchases by Russians in summer 2026.
- CNews, October 2, 2026 — the wave of VPN-related blocks and the May wave of Russian profile blocks.
- TAdviser — additional location-detection parameters and the requirement to be in a supported country.
- Habr: “Claude got banned” — a test of 25 services from a Russian IP on October 1, 2026, including a case of blocked corporate accounts.
- Habr: blocks in Hong Kong — emails about “suspicious activity,” appeals, and the May wave in Russia.
- Habr: tokenizer test — discrepancies in token counts with an intermediary and the tokenization density of Russian text.
- Habr: changes in Gemini from October 9 — reduced free access.
- RSWebsols — a summary of The Information’s report on Meta and Microsoft reducing internal use of Claude.
- CNews, July 9, 2026 — sovereign and national models, and the transition period for systems created before March 1, 2027.
The blocking details are based on user reports and media summaries; Anthropic has not issued an official statement about a new policy. The seven-day plan, option tables, and evaluation criteria are an editorial analysis by AIrassvet; the article does not include the outlet’s own model tests.
How AIrassvet helps reduce dependence on a single AI vendor
If a block has already stopped your work, or you want to prepare in advance, the issue can be broken down step by step. On the AIrassvet website, which also uses the name “Vladimir Lomtev’s AI Studio,” the following applicable service areas are described:
- Auditing processes where AI is used: a registry of dependencies, data, and criticality for each scenario.
- Deploying and fine-tuning an open-source language model on company servers for processes involving sensitive data.
- Integrating Russian and foreign models through a single layer where the vendor is changed by configuration, not by rewriting code.
- Moving instructions, templates, and working documents into a corporate knowledge base with search for an AI assistant.
The first step is to choose one critical process, document its current metrics, data sources, constraints, and the criterion you’ll use to accept a backup model.
Bottom line: access to the model is not a company asset
The October wave of blocks showed what was already true: access to a foreign model from Russia depends on the provider’s terms and can disappear without warning. Companies protect not accounts but processes — through a dependency register, exported context, a unified call layer, and a verified backup model.
Start with an inventory this week. Even if your accounts are still working, in seven days you’ll have a plan in which the next block becomes a configuration switch, not an operational shutdown.