Why Claude Code is not supported in some regions
This gets flattened into "a network problem", so people start optimising nodes and IP reputation. Read the actual rule once and the picture changes: the network is the outermost symptom, not the cause.
Three layers stack. Anthropic maintains a list of supported countries and does not offer the service in the rest; since September 2025 the restriction also covers entities more than 50% owned by companies headquartered in unsupported regions; and national export controls bind on top. The middle layer is decisive because it judges who the account belongs to, not where you connect from.
Layer one: the service-region list
Anthropic publishes a list of countries and regions where it offers its services; mainland China is not on it. Read that as a commercial decision rather than a firewall: the model has no geography, whether you can reach the endpoint is a separate question, and whether you are permitted to use it is a contract question. Those three get mixed together constantly, and that mixing is where most bad decisions start.
The list is revised from time to time, so check the live document rather than any article quoting it.
Layer two: ownership, the decisive one
Anthropic's Supported Regions Policy lists three conditions in parallel, and the first is about the person, not the company: "Use by persons while physically located in an unsupported region, including users and personnel of otherwise supported entities." The September 2025 update added the ownership limbs — entities incorporated or headquartered in an unsupported region, and entities more than 50% owned or controlled by parties there.
Read it closely: it constrains equity structure, not office location and not network path. Consequences follow directly:
- A company registered in Singapore, staffed in Tokyo, but majority-owned from a restricted region is in scope.
- A person in California on a residential connection whose account is judged to belong to such an entity is in scope.
- Conversely, an account in perfect standing still hits the first condition the moment its owner is physically in an unsupported region — a green card and a US company are irrelevant to that clause.
| What people assume | What is actually evaluated |
|---|---|
| Which country the IP sits in | Ownership of the account and the entity behind it |
| Whether a proxy is in use | Registration details, payment instrument, corporate structure |
| Being abroad is enough | Direction matters: leaving an unsupported region is fine, being in one is itself a violation |
| A new node means a fresh start | The ownership conclusion does not move |
Layer three: export controls
The outermost layer is national policy. Export controls on advanced models and compute have tightened, which means availability is no longer purely a vendor decision — it is squeezed between what a company wants to sell and what two governments permit.
The practical implication is the useful part: this layer is entirely outside an individual's influence. No technical trick operates on it, because what you face is not code to be bypassed but a rule that can be rewritten at any time. Knowing that is worth something — it is a variable you can stop spending money on.
Three arguments that do not survive
"A proxy makes me American"
That conflates network location with identity. A proxy changes where packets leave from; it changes neither the registration entity nor the payment instrument. If the verdict rests on the latter, the former is decoration.
"A residential IP makes me safe"
Residential addresses reduce the odds of being classified as datacenter traffic — a false-positive optimisation. The 2026 instrumentation did not read the IP at all: it read the system timezone and the domain in ANTHROPIC_BASE_URL. The thing being paid for was not in the decision path.
"Switching to an open model solves it"
It solves availability — that part stops depending on the region list. It does not solve workflow quality: agent-loop stability, tool-call fidelity and the conventions you accumulated in CLAUDE.md all degrade. Judging that trade is what the migration guide is for.
What "unsupported" costs in practice
- The account can be reclaimed without a restoration promise. Usage history, subscription state and cloud sessions can all go with it.
- Refund treatment is not guaranteed. Partial, full or none, depending on circumstances and payment channel.
- Appeals rarely produce a substantive answer. You can write one; do not build a plan around it.
Together those point at one pragmatic conclusion, independent of which tools you prefer: do not put critical throughput behind an account that can be terminated unilaterally. That is ordinary risk management, not a statement about any vendor.
Is it fine to use Claude Code while abroad?
Three conditions, any one sufficient: physical presence in an unsupported region; an entity incorporated or headquartered there; or majority ownership from there. The first is independent of ownership, so "my company is in the US" is not a defence while you are physically in mainland China.
Does a cleaner IP address help?
Not for the ownership question. A residential exit reduces the chance of being mistaken for datacenter traffic, but the 2026 instrumentation read the system timezone and the base-URL domain rather than the IP. See how bans are detected.
Why does Anthropic apply these restrictions at all?
Three reasons stack: the company does not offer the service in that region, its terms name restricted entities explicitly, and export controls on advanced models constrain sales. Only the first two are commercial choices; the third is outside anyone's control.